NEARLY–LIGHT CYCLES IN EMBEDDED GRAPHS AND CROSSING–CRITICAL GRAPHS
视网膜功能启发的边缘检测层级模型
视网膜功能启发的边缘检测层级模型郑程驰 1范影乐1摘 要 基于视网膜对视觉信息的处理方式, 提出一种视网膜功能启发的边缘检测层级模型. 针对视网膜神经元在周期性光刺激下产生适应的特性, 构建具有自适应阈值的Izhikevich 神经元模型; 模拟光感受器中视锥细胞、视杆细胞对亮度的感知能力, 构建亮度感知编码层; 引入双极细胞对给光−撤光刺激的分离能力, 并结合神经节细胞对运动方向敏感的特性, 构建双通路边缘提取层; 另外根据神经节细胞神经元在多特征调控下延迟激活的现象, 构建具有脉冲延时特性的纹理抑制层; 最后将双通路边缘提取的结果与延时抑制量相融合, 得到最终边缘检测结果. 以150张来自实验室采集和AGAR 数据集中的菌落图像为实验对象对所提方法进行验证, 检测结果的重建图像相似度、边缘置信度、边缘连续性和综合指标分别达到0.9629、0.3111、0.9159和0.7870, 表明所提方法能更有效地进行边缘定位、抑制冗余纹理、保持主体边缘完整性. 本文面向边缘检测任务, 构建了模拟视网膜对视觉信息处理方式的边缘检测模型, 也为后续构建由视觉机制启发的图像计算模型提供了新思路.关键词 边缘检测, 视网膜, Izhikevich 模型, 神经编码, 方向选择性神经节细胞引用格式 郑程驰, 范影乐. 视网膜功能启发的边缘检测层级模型. 自动化学报, 2023, 49(8): 1771−1784DOI 10.16383/j.aas.c220574Multi-layer Edge Detection Model Inspired by Retinal FunctionZHENG Cheng-Chi 1 FAN Ying-Le 1Abstract Based on the processing of visual information by the retina, this paper proposes a multi-layer model of edge detection inspired by retinal functions. Aiming at the adaptive characteristics of retinal neurons under periodic light stimulation, an Izhikevich neuron model with adaptive threshold is established; By simulating the perception ability of cones and rods for luminance and color in photoreceptors, the luminance perception coding layer is con-structed; By introducing the ability of bipolar cells for separating light stimulation, and combining with the charac-teristics of ganglion cells sensitive to the direction of movement, a multi-pathway edge extraction layer is constructed;In addition, according to the phenomenon of delayed activation of ganglion cell neurons under multi-feature regula-tion, a texture inhibition layer with pulse delay characteristics is constructed; Finally, by fusing the result of multi-pathway edge extraction with the delay suppression amount, the final edge detection result is obtained. The 150colony images from laboratory collection and AGAR dataset are used as experimental objects to test the proposed method. The reconstruction image similarity, edge confidence, edge continuity and comprehensive indicators of the detection results are 0.9629, 0.3111, 0.9159 and 0.7870, respectively. The results show that the proposed method can better localize edges, suppress redundant textures, and maintain the integrity of subject edges. This research is oriented to the task of edge detection, constructs an edge detection model that simulates the processing of visual information by the retina, and also provides new ideas for the construction of image computing model inspired by visual mechanism.Key words Edge detection, retina, Izhikevich model, neural coding, direction-selective ganglion cells (DSGCs)Citation Zheng Cheng-Chi, Fan Ying-Le. Multi-layer edge detection model inspired by retinal function. Acta Automatica Sinica , 2023, 49(8): 1771−1784边缘检测作为目标分析和识别等高级视觉任务的前级环节, 在图像处理和工程应用领域中有重要地位. 以Sobel 和Canny 为代表的传统方法大多根据相邻像素间的灰度跃变进行边缘定位, 再设定阈值调整边缘强度和冗余细节[1]. 虽然易于计算且快速, 但无法兼顾弱边缘感知与纹理抑制之间的有效性, 难以满足复杂环境下的应用需要. 随着对生物视觉系统研究的进展, 人们对视觉认知的过程和视觉组织的功能有了更深刻的了解. 许多国内外学者在这些视觉组织宏观作用的基础上, 进一步考虑神经编码方式与神经元之间的相互作用, 并应用于边缘检测中. 这些检测方法大多首先会选择合适的神经元模型模拟视觉组织细胞的群体放电特性, 再关联例如视觉感受野和方向选择性等视觉机制, 以不收稿日期 2022-07-14 录用日期 2022-11-29Manuscript received July 14, 2022; accepted November 29,2022国家自然科学基金(61501154)资助Supported by National Natural Science Foundation of China (61501154)本文责任编委 张道强Recommended by Associate Editor ZHANG Dao-Qiang1. 杭州电子科技大学模式识别与图像处理实验室 杭州 3100181. Laboratory of Pattern Recognition and Image Processing,Hangzhou Dianzi University, Hangzhou 310018第 49 卷 第 8 期自 动 化 学 报Vol. 49, No. 82023 年 8 月ACTA AUTOMATICA SINICAAugust, 2023同的编码方式将输入的图像转化为脉冲信号, 经过多级功能区块处理和传递后提取出图像的边缘. 其中, 频率编码和时间编码是视觉系统编码光刺激的重要方式, 在一些计算模型中被广泛使用. 例如,文献[2]以HH (Hodgkin-Huxley)神经元模型为基础, 使用多方向Gabor滤波器模拟神经元感受野的方向选择性, 实现神经元间连接强度关联边缘方向,将每个神经元的脉冲发放频率作为边缘检测的结果输出, 实验结果表明其比传统方法更有效; 文献[3]在 LIF (Leaky integrate-and-fire) 神经元模型的基础上进行改进, 引入根据神经元响应对外界输入进行调整的权值, 在编码的过程中将空间的脉冲发放转化为时序上的激励强度, 实现强弱边缘分类, 对梯度变化幅度小的弱边缘具有良好的检测能力. 除此之外, 也有关注神经元突触间的相互作用, 通过引入使突触的连接权值产生自适应调节的机制来提取边缘信息的计算方法. 例如, 文献 [4] 构建具有STDP (Spike-timing-dependent plasticity) 性质的神经元模型, 根据突触前后神经元首次脉冲发放时间顺序来增强或减弱突触连接, 对真伪边缘具有较强的辨别能力; 文献 [5] 则在构建神经元模型时考虑了具有时间不对称性的STDP机制, 再融合方向特征和侧抑制机制重建图像的主要边缘信息, 其计算过程对神经元突触间的动态特性描述更加准确.更进一步, 神经编码也被应用于实际的工程需要.例如, 文献 [6]针对现有的红外图像边缘检测算法中存在的缺陷, 构建一种新式的脉冲神经网络, 增强了对红外图像中弱边缘的感知; 文献 [7] 则通过模拟视皮层的处理机制, 使用包含左侧、右侧和前向3条并行处理支路的脉冲神经网络模型提取脑核磁共振图像的边缘, 并将提取的结果用于异常检测,同样具有较好的效果. 上述方法都在一定程度上考虑了视觉组织中神经元的编码特性以及视觉机制,与传统方法相比, 在对复杂环境的适应性更强的同时也有较高的计算效率. 但这些方法都未能考虑到神经元自身也会随着外界刺激产生适应, 从而使活动特性发生改变. 此外, 上述方法大多也只选择了频率编码、时间编码等编码方式中的一种, 并不能完整地体现视觉组织中多种编码方式的共同作用.事实上, 在对神经生理实验和理论的持续探索中发现, 视觉组织(以视网膜为例)在对视觉刺激的加工中就存在着丰富的动态特性和编码机制[8−9]. 视网膜作为视觉系统中的初级组织结构, 由多种不同类型的细胞构成, 共同组成一个纵横相连、具有层级结构的复杂网络, 能够针对不同类型的刺激性选择相应的编码方式进行有效处理. 因此, 本文面向图像的边缘检测任务, 以菌落图像处理为例, 模拟视网膜中各成分对视觉信息的处理方式, 构建基于视网膜动态编码机制的多层边缘检测模型, 以适应具有多种形态结构差异的菌落图像边缘检测任务.1 材料和方法本文提出的算法流程如图1所示. 首先, 根据视网膜神经元在周期性光刺激下脉冲发放频率发生改变的特性, 构建具有自适应阈值特性的Izhikevich 神经元模型, 改善神经元的同步发放能力; 其次, 考虑光感受器对强弱光和颜色信息的不同处理方式编码亮度信息, 实现不同亮度水平目标与背景的区分;然后, 引入固视微动机制, 结合神经节细胞的方向选择性和给光−撤光通路的传递特性, 将首发脉冲时间编码的结果作为双通路的初级边缘响应输出;随后, 模拟神经节细胞的延迟发放特性, 融入对比度和突触前后偏好方向差异, 计算各神经元的延时抑制量, 对双通路的计算结果进行纹理抑制; 最后,整合双通路边缘信息, 将二者融合为最终的边缘检测结果.1.1 亮度感知编码层构建神经元模型时, 本文综合考虑对神经元生理特性模拟的合理性和进行仿真计算的高效性, 以Izhikevich模型[10]为基础构建神经元模型. Izhike-vich模型由Izhikevich在HH模型的基础上简化而来, 在保留原模型对神经元放电模式描述的准确性的同时, 也具有较低的时间复杂度, 适合神经元群体计算时应用, 其表达式如下式所示v thv th 其中, v为神经元的膜电位, 其初始值设置为 −70; u为细胞膜恢复变量, 设置为14; I为接收的图像亮度刺激; 为神经元脉冲发放的阈值, 设置为30; a描述恢复变量u的时间尺度, b描述恢复变量u 对膜电位在阈值下波动的敏感性, c和d分别描述产生脉冲发放后膜电位v的重置值和恢复变量u的增加程度, a, b, c, d这4个模型参数的典型值分别为0.02、0.2、−65和6. 若某时刻膜电位v达到,则进行一次脉冲发放, 同时该神经元对应的v被重置为c, u被重置为u + d.适应是神经系统中广泛存在的现象, 具体表现为神经元会根据外界的刺激不断地调节自身的性质. 其中, 视网膜能够适应昼夜环境中万亿倍范围的光照变化, 这种适应能够帮助其在避免饱和的同时保持对光照的敏感性[11]. 研究表明, 视网膜持续1772自 动 化 学 报49 卷接受外界周期性光刺激时, 光感受器会使神经元细胞的活动特性发生改变, 导致单个神经元的发放阈值上升, 放电频率下降; 没有脉冲发放时, 对应阈值又会以指数形式衰减, 同时放电频率逐渐恢复[12].因此, 本文在Izhikevich 模型的基础上作出改进,加入根据脉冲发放频率对阈值进行自适应调节的机制, 如下式所示τ1τ2τ1τ2v th τ1v th τ2其中, 和 分别为脉冲发放和未发放时阈值变化的时间常数, 其值越小, 阈值变化的幅度越大, 神经元敏感性变化的过程越快; 反之, 则表示阈值变化的幅度越小, 神经元敏感性变化的过程也就越慢.生理学实验表明, 在外界持续光刺激下, 神经元对刺激产生适应导致放电频率降低后, 这种适应衰退的过程比产生适应的过程通常要长数倍[13]. 因此,为了在准确模拟生理特性的同时保证计算模型的性能, 本文将 和 分别设置为20和40. 这样, 当某时刻某个神经元产生脉冲发放时, 则对应阈值 根据 的值升高, 神经元产生适应, 活跃度降低; 反之, 对应阈值 根据 的值下降, 神经元的适应衰退, 活跃度提升. 实现限制活跃神经元的脉冲发放频率, 促进不活跃神经元的脉冲发放, 改善神经元群体的同步发放能力, 减少检测目标内部冗余. 图2边缘检测结果图 1 边缘检测算法原理图Fig. 1 Principle of edge detection algorithm8 期郑程驰等: 视网膜功能启发的边缘检测层级模型1773显示了改进前后的Izhikevich 模型对图像进行处理后目标内部冗余情况.0∼255为了规范检测目标图像的亮度范围, 本文将输入的彩色图像Img 各通路的亮度映射到 区间内, 如下式所示Img (;i )I (;i )其中, 和 表示经亮度映射前和映射后的R 、G 、B 三种颜色分量图像; max(·) 和min(·)分别计算对应分量图像中的最大和最小像素值.光感受器分两类, 分别为视锥细胞和视杆细胞[14], 都能将接收到的视觉刺激转化为电信号, 实现信息的编码和传递. 其中, 视锥细胞能够根据外界光刺激的波长来分解为三个不同的颜色通道[15].考虑到人眼对颜色信息的敏感性能有效区分离散目标与背景, 令图像中的每个像素点对应一个神经元,将R 、G 、B 三种颜色分量图像分别输入上文构建的神经元模型中, 在一定时间范围内进行脉冲发放,如下式所示fires (x,y ;i )其中, 为每个神经元的脉冲发放次数,函数Izhikevich(·)表示式(2)给出的神经元模型.视杆细胞对光线敏感, 主要负责弱光环境下的外界刺激感知. 当光刺激足够强时, 视杆细胞的感知能力达到饱和, 视觉系统转为使用视锥细胞负责亮度信息的处理[16]. 因此, 除了对颜色信息敏感外,视锥细胞对强光也有高度辨别能力. 考虑到作为检测对象的图像中, 目标与背景具有不同的亮度水平,本文构建一种综合视锥细胞和视杆细胞亮度感知能力的编码方法, 以适应目标与背景不同亮度对比的多种情况, 如下式所示I base I base (x,y )fires Res (x,y )其中, var(·) 计算图像亮度方差; ave(·) 计算图像亮度均值. 本文取三种颜色分量图像中方差最大的一幅作为基准图像 , 对于其中的像素值 ,将其中亮度低于平均亮度的部分设置为三种颜色分量脉冲发放结果的最小值, 反之设置为最大值, 最终得到模型的亮度编码结果 , 实现在图像局部亮度相对较低的区域由视杆细胞进行弱光感知, 亮度较高区域由视锥细胞处理, 强化计算模型对不同亮度目标和背景的区分能力, 凸显具有弱边缘的对象. 图3显示了亮度感知编码对存在弱边缘的对象的感知能力.1.2 基于固视微动的多方向双通路边缘提取层Img gray 人眼注视目标时, 接收的图像并非是静止的,而是眼球以每秒2至3次的微动使投射在视网膜上的图像发生持续运动, 不断地改变照射在光感受器上的光刺激[17]. 本文考虑人眼的固视微动机制,在原图像的灰度图像 上构建大小为3×3的微动作用窗口temp , 使窗口接收到的亮度信息朝8个方向进行微动, 如下式所示p i q i θi temp θi d x d y 其中, 和 是用于决定微动方向 的参数, 其值被设置为 −1、0或1, 通过计算反正切函数能够得到以45° 为单位、从0° 到315° 的8个角度的微动方向, 对应8个微动结果窗口 ; 和 分别表示水平和竖直方向的微动尺度; Dir 为计算得到(a) 原图(a) Original image (b) Izhikevich 模型(b) Izhikevich model (c) 改进的 Izhikevich 模型(c) Improved Izhikevich model图 2 改进前后的Izhikevich 模型对图像进行脉冲发放的结果对比图Fig. 2 Comparison of the image processing results of the Izhikevich model before and after improvement1774自 动 化 学 报49 卷Dir (x,y )的微动方向矩阵, 其中每个像素点的值为 ;sum(·) 计算窗口中像素值的和. 本文取每个微动窗口前后差异最大的方向作为该点的偏好方向, 分别用数字1 ~ 8表示.视网膜存在一类负责对运动刺激编码、具有方向选择性的神经节细胞 (Direction-selective gangli-on cells, DSGCs)[18]. 经过光感受器处理, 转化为电信号的视觉信息, 通过双极细胞处理后传递给神经节细胞. 双极细胞可分为由光照增强 (ON) 激发的细胞和由光照减弱 (OFF) 激发的细胞[19], 分别将信号输入给光通路 (ON-pathway)和撤光通路 (OFF-pathways) 两条并行通路[20], 传递给光运动和撤光运动产生的刺激. 而神经节细胞同样包括ON 和OFF 两种, 会对给光和撤光所产生的运动方向做出反应[21]. 因此, 本文构造5×5大小的对特定方向微动敏感的神经节细胞感受野窗口, 将其对偏好方向和反方向微动所产生的响应分别作为给光通路和撤光通路的输入. 以偏好方向为45° 的方向选择性神θi fires Res S xy ∗通过上述定义, 可以形成以45° 为单位、从0°到315° 的8个方向的感受野窗口, 与上文 的8个方向对应. 之后本文在亮度编码结果 上构筑与感受野相同大小的局部窗口 , 根据最优方向矩阵Dir 对应窗口中心点的方向, 取与其相同和相反方向的感受野窗口和亮度编码结果进行卷积运算 (本文用符号 表示卷积运算), 分别作为ON 和OFF 通道的输入, 如下式所示T ON T OFF 考虑到眼球微动能够将静止的空间场景转变为视网膜上的时间信息流, 激活视网膜神经元的发放,同时ON 和OFF 两通路也只在光刺激的呈现和撤去的瞬时产生电位发放, 因此本文采用首发脉冲时间作为编码方式, 将 和 定义为两通路首次脉冲发放时间构成的时间矩阵, 并作为初级边缘响应的结果. 将1个单位的发放时间设置为0.25, 当总发放时间大于30时停止计算, 此时还未进行发放的神经元即被判断为非边缘.1.3 多特征脉冲延时纹理抑制层视网膜神经节细胞在对光刺激编码的过程中,外界刺激特征的变化会显著影响神经元的反应时间. 研究发现, 当刺激对比度增大时, 神经元反应延时会减小, 更快速地进行脉冲发放; 反之, 则反应延时增大, 抑制神经元的活性[22]. 除此之外, 方向差异也会影响神经元活动, 突触前后偏好方向相似的神经元更倾向于优先连接, 在受到外界刺激时能够更快被同步激活[23]. 因此, 本文引入视网膜的神经元延时发放机制, 考虑方向和对比度对神经元敏感性的影响, 构造脉冲延时抑制模型. 首先结合局部窗口权重函数计算图像对比度, 如下式所示ω(x i ,y i )其中, 为窗口权重函数, L 为亮度图像, Con(a) 原图(a) Original image (b) Izhikevich 模型(b) Izhikevich model (c) 改进的 Izhikevich 模型(c) Improved Izhikevich model (d) 亮度感知编码(d) Luminance perception coding图 3 不同方式对存在弱边缘的菌落图像的处理结果Fig. 3 Different ways to process the image of colonies with weak edges8 期郑程驰等: 视网膜功能启发的边缘检测层级模型1775S xy x i y i µ=∑x i ,y i ∈S xy ω(x i ,y i )为对比度图像, 为以(x , y )为中心的局部窗口,( , ) 为方窗中除中心外的周边像素, ws 为局部方窗的窗长, . 之后考虑局部方窗中心神经元和周边神经元方向差异, 同时用高斯函数模拟对比度大小与延时作用强度之间的关系, 构建脉冲延时抑制模型, 如下式所示D Dir (x,y )D Con (x,y )D (x,y )∆Dir (x i ,y i )min {|θ(x i ,y i )−θ(x,y )|,2π−|θ(x i ,y i )−θ(x,y )|}δ其中, 和 分别表示方向延时抑制量和对比度延时抑制量; 为计算得到的综合延时抑制量; 为突触前后神经元微动方向的差异, 被定义为 ; 用于调节对比度延时抑制量.T ON T OFFRes ON Res OFF 将上文计算得到的两个时间矩阵 和 中进行过脉冲发放的神经元与综合延时抑制量相加, 同样设置1个单位的发放时间为0.25, 将经延时作用后总发放时间大于30的神经元设置为不发放, 即判定为非边缘, 反之则判定为边缘. 根据式(19)和式(20) 得到两通道边缘检测结果 和. 最后, 将两通道得到的结果融合, 得到最终边缘响应结果Res ,如下式所示2 算法流程基于视网膜对视觉信息的处理顺序和编码特性, 本文构建图4所示的算法流程, 具体步骤如下:1) 根据视网膜在外界持续周期性光刺激下产生的适应现象, 在式(1)所示的Izhikevich 模型上作出改进, 构建如式(2)所示的具有自适应阈值的Izhikevich 模型.2) 根据式(3)将作为检测目标的图像映射到0 ~ 255区间规范亮度范围, 接着分离3种通道的颜色分量, 根据式(4)输入到改进的Izhikevich 模型中进行脉冲发放.3) 根据式(5)的方差计算提取出基准图像, 再结合基准图像根据式(6)对三通道脉冲发放的结果进行亮度感知编码, 得到亮度编码结果.4) 考虑人眼的固视微动机制, 根据式(7)和式(8)通过原图的灰度图像提取每个神经元的偏好方向, 得到微动方向矩阵, 接着根据式(9)和式(10)构筑8个方向的方向选择性神经节细胞感受野窗口.5) 根据式(11)和式(12), 将感受野窗口与亮度编码图像作卷积运算, 并输入Izhikevich 模型中得到ON 和OFF 通路的首发脉冲时间矩阵, 作为两通道的初级边缘响应.6) 根据式(13) ~ 式 (15), 结合局部窗口权重计算图像对比度.7) 考虑对比度和突触前后偏好方向对脉冲发放的延时作用, 根据式(16) ~ 式 (18)构建延时纹理抑制模型, 并根据式(19)和式(20)将纹理抑制模型和两通道的初级边缘响应相融合.8) 根据式(21)将两通路纹理抑制后的结果在神经节细胞处进行整合, 得到最终边缘响应结果.3 结果为了验证本文方法用于菌落边缘检测的有效性, 本文选择Canny 方法和其他3种同样基于神经元编码的边缘检测方法作为横向对比, 并进行定性、定量分析. 首先, 选择文献[4]提出的基于神经元突触可塑性的边缘检测方法(Synaptic plasticity model, SPM), 用于对比本文方法对弱边缘的增强效果; 其次, 选择文献[24]提出的基于抑制性突触的多层神经元群放电编码的边缘检测方法 (Inhibit-ory synapse model, ISM), 验证本文的延时抑制层在抑制冗余纹理方面的有效性; 然后, 选择文献[25]提出的基于突触连接视通路方向敏感的分级边缘检测方法(Orientation sensitivity model, OSM), 对比本文方法在抑制冗余纹理的同时保持边缘提取完整性上的优势; 最后, 还以本文方法为基础, 选择去除亮度感知编码后的方法(No luminance coding,NLC)作为消融实验, 以验证本文方法模拟光感受器功能的亮度感知编码模块的有效性.本文使用实验室在微生物学实验中采集的菌落图像和AGAR 数据集[26]作为实验对象. 前者具有丰富的颜色和形态结构, 用于检验算法对复杂检测环境的适应性; 后者则存在更多层次强度的边缘信息, 菌落本身与背景的颜色和亮度水平也较为相近,用于检测算法对颜色、亮度特征和弱边缘的敏感性.本文通过局部采样生成150张512×512像素大小的测试图像, 其中38张来自实验室采集, 112张来自AGAR 数据集. 然后分别使用上文的6种边缘1776自 动 化 学 报49 卷检测算法提取图像边缘, 使每种算法得到150张边缘检测结果, 其中部分检测结果如图5所示.定性分析图5可知, Canny 、SPM 和ISM 方法在Colony4和Colony5等存在弱边缘的图像中往往会出现大面积的边缘丢失. OSM 方法对弱边缘的敏感性强于以上3种方法, 但仍然会出现不同程度的边缘断裂, 且在调整阈值时难以均衡边缘连续性和目标菌落内部冗余. NLC 方法同样丢失了Colony4和Colony5中几乎所有的边缘, 对于Colony3也只能检出其中亮度较低的菌落内部, 对于梯度变化不明显的边缘辨别力差. 与其他方法相比, 本文方法检出的边缘更加显著且完整性更高, 对于弱边缘也有很强的检测能力, 在Colony3、Colony4和Colony5等存在多层次水平强弱边缘的菌落图像中能够取得较好的检测结果. 为了对检测结果进行定量分析并客观评价各方法的优劣, 计算边缘图像重建相似度MSSIM [27]对检测结果进行重建, 并计算重建图像与原图像的相似度作为边缘定位的准确性RGfires (R)fires (G)亮度编码结果Luminance codingresult方差计算Variance1 2 3ON-result对比度Contrast脉冲延时抑制量Neuron spiking delay感受野窗口感受野窗口DSGC templateOFF-通路输出OFF-result 5)6)7)图 4 边缘检测算法流程图Fig. 4 The procedure of edge detection algorithm8 期郑程驰等: 视网膜功能启发的边缘检测层级模型1777图 5 Colony1 ~ Colony5的边缘检测结果(第1行为原图; 第2行为Canny 检测的结果; 第3行为SPM 检测的结果; 第4行为ISM 检测的结果; 第5行为OSM 检测的结果; 第6行为NLC 检测的结果; 第7行为本文方法检测的结果)Fig. 5 Edge detection results of Colony1 to Colony5 (The first line is original images; The second line is the results of Canny; The third line is the results of SPM; The fourth line is the results of ISM; The fifth line is the results of OSM;The sixth line is the results of NLC; The seventh line is the results of the proposed method)1778自 动 化 学 报49 卷指标. 首先对检测出的边缘图像做膨胀处理, 之后将原图像上的像素值赋给膨胀后边缘的对应位置,得到的图像记为ET , 则边缘重建如下式所示T k ET d k 其中, 为图像 上3×3窗口中8个方向的周边像素, 为窗口中心像素点与周边像素的距离, 计算得到重建图像R . 重建图像的相似度指标如下式所示µA µB σA σB σAB 其中, 和 为原图像和重建图像的灰度均值, 和 为其各自的标准差, 为原图像与重建图像之间的协方差. 将原图像和重建图像各自分为N 个子图, 并分别计算相似度指标SSIM , 得到平均相似度指标MSSIM . 除此之外, 为了验证边缘检测方法检出边缘的真实性和对菌落内部冗余纹理的抑制能力, 本文计算边缘置信度BIdx [28], 根据边缘两侧灰度值的跃变程度判断边缘的真伪. 边缘置信度指标如下式所示σij E (x i k ,y ik )(x i ,y i )d i其中, 为边缘像素在原图像对应位置的邻域标准差, EdgeNum 为边缘像素数量. 另外, 本文进一步计算边缘连续性 CIdx [29]来验证检出目标的边缘完整性. 首先将得到的边缘图像E 分割为m 个区域, 分别计算每个区域中的边缘像素 到其空间中心 的距离 ,则连续性指标如下式所示c i k C i n i 其中, 为边缘连续性的贡献值, D 为阈值, 为第i 个区域的像素点的连续性贡献值之和,为第i 个区域边缘像素点数量. 最后, 将计算得到的3个指标根据下式融合, 得到综合评价指标EIdx [21]其中, row 和col 分别为原图像的行数和列数. 于是, 检测图像的各项性能指标如表1 ~ 表5所示, 图像重建的结果如图6所示.表 1 不同检测方法下的重建相似度MSSIM Table 1 MSSIM of different methodsSerial number MSSIMCanny SPMISMOSMNLC本文方法Colony10.74520.77250.83570.92650.91750.9371Colony20.79510.79710.84900.95280.94470.9725Colony30.85760.86620.83140.91490.83370.9278Colony40.96900.98270.98380.98870.98930.9972Colony50.96340.97580.97800.97710.98830.9933表 2 不同检测方法下的边缘置信度BIdx Table 2 BIdx of different methodsSerial number BIdxCanny SPMISMOSMNLC本文方法Colony10.49880.46180.43070.58010.50580.6026Colony20.18210.15370.15530.33650.46150.4479Colony30.19830.15100.16100.26340.12630.3257Colony40.16310.14880.19060.14370.15210.2016Colony50.16200.18960.19020.18820.17350.1654表 3 不同检测方法下的边缘连续性CIdxTable 3 CIdx of different methodsSerial numberCIdxCanny SPMISMOSMNLC本文方法Colony10.83770.85300.86010.86760.97490.9652Colony20.80690.86550.85330.82930.91770.9518Colony30.80640.74080.72930.82690.77640.9406Colony40.81430.86110.90440.84300.90150.9776Colony50.90470.84480.86320.85920.87090.95718 期郑程驰等: 视网膜功能启发的边缘检测层级模型1779。
光照变化 brdf 高光谱 反射率重建
光照变化brdf 高光谱反射率重建
光照变化(Lighting Variation)是指在不同光源照射下,物体表面产生的亮度和颜色的变化。
光照变化可以由不同的光源强度、方向和色温引起。
在计算机图形学和计算机视觉领域,研究光照变化可以帮助改善渲染、识别和重建等任务的表现。
BRDF (Bidirectional Reflectance Distribution Function) 是描述表面反射性质的函数。
它定义了在给定入射光方向和出射光方向下,表面单位面积上反射光的分布情况。
BRDF 描述了光线与表面之间的相互作用,影响着物体在不同光照条件下的外观。
高光谱(Hyperspectral)是指在非常细微的光谱范围内获取大量连续的光谱数据。
传统的彩色图像通常只包含可见光频段的红、绿、蓝三个通道,而高光谱图像则可以包含从紫外线到红外线的更广泛的频段。
利用高光谱图像,可以更详细地观察和分析物体的光谱特征,从而获得更多关于物体材质和组成的信息。
反射率重建(Reflectance Reconstruction)是指根据观测到的光照变化和物体表面的光反射特性,估计或还原出物体各个点的真实反射率值。
这一过程可以利用物体的BRDF模型和高光谱图像的数据进行。
通过反射率重建,我们可以获得更加准确和细致的物体表面光学特性的估计结果,有助于计算机图形渲染、三维重建、材料识别等应用。
翻译1
一种新的制备Sn-C/LiFe0.1Co0.9PO4液态的锂离子全电池的方法。
摘要:在这项工作中,我们所提出的锂离子电池使用的阴极是以Fe取代LiCoPO4做为涂层的C,阳极是纳米结构的Sn-C,电解液是以N-丁基-N-甲基吡咯-烷氧基锂-双三氟甲烷磺酰亚胺锂不易燃材料为基础介质。
初步结果表明,电池电压约为4.5 V,由于电化学嵌入/脱嵌过程取决于Co3+/Co2+氧化还原电对在阴极的反应和Li-Sn-C在阳极的合金化与去合金化过程,所以容量与特定的容量接近,为90mAh g-1。
正文:由于LiCoPO4橄榄石高电势(例如Li/Li+,4.8V)和大的理论容量(约为167mAh g-1),所以他的出现使阴极锂离子电池有了新的前途。
然而低的电子电导率和锂离子迁移率低的晶格形式仍然限制LiCoPO4实际应用[2 - 4]。
碳涂层和金属掺杂已经被提议作为合适的方法来提高橄榄石阴极的电子和离子电导率[5]。
特别是,添加小数量的Fe的LiCoPO4会大大改善锂电池的电化学性能[6、7]。
传统电解质低的电化学稳定仍然是阻碍阴极高压工作的关键。
事实上,在高压电极工作时这种电解质的分解会导致容量的快速下降。
特别是,氟化物杂质的存在,例如微量的HF,比如在使用LiCoPO4电极时,以LiPF6为基础成分电解质可以诱导电池的进一步衰退,这是由于F 阴离子的亲核攻击使得橄榄石上的P原子脱锂(带电)[8]。
这些问题中,除了由于易挥发和易燃有机烷基、碳酸盐岩存在于常见的电解质而导致安全问题可能成功地制约了使用高度稳定之外,电解液中存在的氟离子也会决定电池的电化学稳定性。
但是,非液态电解质的使用可能在常见的石墨阳极锂离子电池使用中引起一些问题,比如结构分层不稳定的固态电解质界面(S EI)的薄膜,从而导致电池不能工作[12]。
这个问题能否有效解决决定新一代的阳极的使用,例如以高容量和非凡的稳定性为表征的锂金属合金化合物等。
上海研制出仿生太阳能电池 光电转化率接近世界最高水平
●辛 旺
的世界最高水平 。
项 目负责 人 、 华 东 师 大 纳 光 电集 成
与先进装备教 育部工 程研 究 中心 主任
孙卓教授展示 了新 型太 阳能 电池 的“ 三
明治 ” 结 构— — 中空 玻 璃 夹 着 一 层 纳米
“ 夹 心 ” 光 , 电转 化 的 玄 机 就 藏 在 这 层 几 十微 米厚 的复合 薄膜 中 。 纳米 “ 夹 心 ” 的
“ 配 方 ” 十 分 独 特 :染 料 充 当 “ 捕 光 手 ” ,
纳米二 氧化钛则是 “ 光 电转换器 ” 。 为 了
让 染料 尽 可 能多“ 吃 ” 太 阳光 ,科研 人 员
还 别 出心 裁地 撒 了点
佐 料 “
”
—
—
种 一
由纳米荧光材料制成 的量子 点 ,让 不 同
波 长 的 阳光 都 能对 上 “ 捕 光 手 ” 的 “ 胃
电池产量 为 9 2 万 千瓦 , 比上 年减少 了
1 1.3 % 。 欧洲 的产量则 同 比大 幅上 升
,达 到
10 6
万千瓦。
美 国的产量为
2 7 万 千 瓦 , 中国大 陆 、 台湾 、 印度等 国家
和地 区 的产量也 急速 上 升 。
随机森林算法改进综述
随机森林算法改进综述发布时间:2021-01-13T10:23:33.577Z 来源:《科学与技术》2020年第27期作者:张可昂[导读] 随机森林是当前一种常用的机器学习算法,张可昂云南财经大学国际工商学院云南昆明 650221摘要:随机森林是当前一种常用的机器学习算法,其是Bagging算法和决策树算法的一种结合。
本文就基于随机森林的相关性质及其原理,对它的改进发展过程给予了讨论。
1、引言当前,随机森林算法得到了快速的发展,并应用于各个领域。
随着研究环境等的变化,且基于随机森林良好的可改进性,学者们对随机森林的算法改进越来越多。
2、随机森林的原理随机森林是一种集成的学习模型,它通过对样本集进行随机取样,同时对于属性也随机选取,构建大量决策树,然后对每一棵决策树进行训练,在决策树中得到许多个结果,最后对所有的决策树的结果进行投票选择最终的结果。
3、随机森林算法改进随机森林的算法最早由Breiman[1]提出,其是由未经修剪树的集合,而这些树是通过随机特征选择并组成训练集而形成的,最终通过汇总投票进行预测。
随机森林的应用范围很广,其可以用来降水量预测[2]、气温预测[3]、价格预测[4]、故障诊断[5]等许多方面。
但是,根据研究对象、数据等不同,随机森林也有许多改进。
例如为了解决在高维数据中很大一部分特征往往不能说明对象的类别的问题,Ye et al.提出了一种分层随机森林来为具有高维数据的随机森林选择特征子空间[6]。
Wang为了解决对高位数据进行分类的问题,提出了一种基于子空间特征采样方法和特征值搜索的新随机森林方法,可以显著降低预测的误差[7]。
尤东方等在研究存在混杂因素时高维数据中随机森林时,实验得出基于广义线性模型残差的方法能有效校正混杂效应[8]。
并且许多学者为了处理不平衡数据问题,对随机森林算法进行了一系列的改进。
为了解决在特征维度高且不平衡的数据下,随机森林的分类效果会大打折扣的问题,王诚和高蕊结合权重排序和递归特征筛选的思想提出了一种改进的随机森林算法,其可以有效的对特征子集进行精简,减轻了冗余特征的影响[9]。
盘点:2018材料领域发表的Nature论文!
盘点:2018材料领域发表的Nature论文!•《Nature》重大突破:复旦大学量子霍尔领域新发现!量子霍尔效应是20世纪以来凝聚态物理领域最重要的科学发现之一,迄今已有四个诺贝尔奖与其直接相关。
但一百多年来,科学家们对量子霍尔效应的研究仍停留于二维体系,从未涉足三维领域。
复旦大学物理学系修发贤课题组首先在该领域实现重大突破,在迈出了从二维到三维的关键一步。
•《Nature》超显微镜观察到锂离子在双层石墨烯中迁移!科学家使用超显微镜,观察到以原子分辨率显示的锂离子在电化学充放电过程中的表现,证明了在单个纳米电池中双层石墨烯发生的可逆锂离子吸收。
实验结果让研究人员感到吃惊,传统的石墨基电池只有少数紧密堆积的锂在两层碳层之间,而在石墨烯纳米电池里发现非常密集的锂层。
•重大突破:吉林大学时隔7年再发《Nature》!团队制备的OLED最大EQE分别达到27%和17%,已接近100%IQE的理论极限值,是目前为止已报道的深红光/近红外光发光二极管(LED)中的最高值。
该研究成果是OLED研究领域的重大突破,展现了发光自由基在有机光电领域的应用前景,为OLED的研究开辟了新的方向。
•石墨烯超导重大发现!中科大少年班校友Nature连发两文!曹原所在团队在魔角扭曲的双层石墨烯中发现新的电子态,可以简单实现绝缘体到超导体的转变,打开了非常规超导体研究的大门。
Nature杂志在2018年3月5日以背靠背的长文形式,在网站刊登了这项还没来得及排版的重大研究成果,并配以评述。
•突破!华侨大学第一篇《Nature》此次论文的刊发标志着魏展画教授团队在钙钛矿电致发光领域取得了重大研究进展。
论文中,他们提出了一种全新的薄膜制备策略并优化了LED器件结构,制备出了高亮度、高量子转换效率和较好稳定性的钙钛矿LED器件。
其中,该钙钛矿LED器件的外量子效率高达20.3%,刷新了世界纪录。
•北科大吕昭平又发《Nature》!同时提高强度和塑性吕昭平教授团队打破人们对传统间隙固溶强化的认知,发现间隙原子的添加不仅能提高合金的强度,也能大幅度提高合金的塑性,并提出了一种设计高强度高塑性金属材料的新的合金设计思路。
基于Retinex理论的低光图像增强算法
第40卷第6期Vol.40㊀No.6重庆工商大学学报(自然科学版)J Chongqing Technol &Business Univ(Nat Sci Ed)2023年12月Dec.2023基于Retinex 理论的低光图像增强算法史宇飞,赵佰亭安徽理工大学电气与信息工程学院,安徽淮南232001摘㊀要:为了解决低光照图像存在的对比度低㊁噪声大等问题,提出一种基于Retinex 理论的卷积神经网络增强模型(Retinex-RANet )㊂它包括分解网络㊁降噪网络和亮度调整网络3部分:在分解网络中融入残差模块(RB )和跳跃连接,通过跳跃连接将第一个卷积层提取的特征与每一个RB 提取的特征融合,以确保图像特征的完整提取,从而得到更准确的反射分量和光照分量;降噪网络以U-Net 网络为基础,同时加入了空洞卷积和注意力机制,空洞卷积能提取更多的图像相关信息,注意力机制可以更好地去除反射分量中噪声,还原细节;亮度调整网络由卷积层和Sigmoid 层组成,用来提高光照分量的对比度;最后将降噪网络去噪后的反射分量和亮度调整网络增强后的光照分量融合,得到最终的增强结果㊂实验结果显示:Retinex-RANet 在主观视觉上不仅提高了低光图像的亮度,还提高了色彩深度和对比度,在客观评价指标上,相较于R2RNet ,PSNR 值上升了4.4%,SSIM 值上升了6.1%㊂结果表明:Retinex-RANet 具有更好的低光图像增强效果㊂关键词:低光增强;残差模块;注意力机制;Retinex 理论中图分类号:TP391㊀㊀文献标识码:A ㊀㊀doi:10.16055/j.issn.1672-058X.2023.0006.008㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀收稿日期:2022-06-09㊀修回日期:2022-07-20㊀文章编号:1672-058X(2023)06-0061-07基金项目:国家自然科学基金面上项目(52174141);安徽省重点研究与开发计划资助项目(202004A07020043);安徽省自然科学基金面上项目(2108085ME158);安徽高校协同创新项目(GXXT -2020-54).作者简介:史宇飞(1997 ),男,安徽安庆人,硕士研究生,从事图像处理研究.通讯作者:赵佰亭(1981 ),男,山东枣庄人,教授,博士,从事图像处理㊁智能控制研究.Email:btzhao@.引用格式:史宇飞,赵佰亭.基于Retinex 理论的低光图像增强算法[J].重庆工商大学学报(自然科学版),2023,40(6):61 67.SHI Yufei ZHAO Baiting.Low-light Image enhancement algorithm based on Retinex theory J .Journal of Chongqing Technology and Business University Natural Science Edition 2023 40 6 61 67.Low-light Image Enhancement Algorithm Based on Retinex Theory SHI Yufei ZHAO BaitingSchool of Electrical and Information Engineering Anhui University of Science and Technology Anhui Huainan 232001 ChinaAbstract In order to solve the problems of low contrast and high noise of low-light images a convolutional neural network enhancement model based on Retinex theory Retinex-RANet is proposed.It includes three parts the decomposition network the noise reduction network and the brightness adjustment network.The residual module RB and the jump connection were incorporated into the decomposition network and the features extracted by the first convolutional layer were fused with each RB extracted feature through the jump connection to ensure the complete extraction of the image features resulting in more accurate reflection and illumination components.The noise reduction network was based on the U-Net network and the cavity convolution and attention mechanism were added at the same time so as to extract more image-related information.The attention mechanism can better remove the noise in the reflected component and restore the details.The brightness adjustment network consists of a convolutional layer and a Sigmoid layer which is used to increase the contrast of the light components.Finally the reflection component after denoising by the noise reduction network and the light component after the brightness adjustment network were fused to obtain the final enhancement result.Experimental results show that Retinex-RANet not only improves the brightness of low-light images in subjective vision but also improves the color depth and contrast.In terms of objective evaluation indicators compared with R2RNet the PSNR value increased by 4.4% and the SSIM value increased by 6.1%.The results show that Retinex-RANet has better重庆工商大学学报(自然科学版)第40卷low-light image enhancement.Keywords low light enhancement residuals module attention mechanism Retinex theory1㊀引㊀言在光照不足㊁不均或者有阴影遮挡等条件下,采集的图像一般都存在噪声过多和对比度弱等问题,而这些问题不但会对图像的品质产生负面影响,还会妨碍一些机器视觉任务的进行㊂对低光照图像进行增强,有助于提高高级视觉性能,如图像识别㊁语义分割㊁目标检测等;也可以在一些实际应用中提高智能系统的性能,如视觉导航㊁自动驾驶等㊂因此,对低光图像增强进行研究是十分必要的㊂低光图像增强方法可分为以下4类:基于直方图均衡化的方法[1],其核心是通过改变图像部分区域的直方图来达到整体对比度提高的效果㊂此类方法可以起到提高图像对比度的作用,但是大多数不够灵活,部分区域仍会出现曝光不足和放大噪音等不好的视觉效果㊂基于去雾的方法[2-3],如一些研究人员[4]利用有雾图像和低光图像之间的相似性,通过已有的去雾算法来增强低光照图像㊂尽管此类方法得到了较好结果,但此类模型的物理解释不够充分,同时对增强后的图像进行去噪可能会导致图像细节模糊㊂基于Retinex理论[5]的方法,其将低光图像分解为光照和反射率两部分,在保持反射率一致性的前提下,增加光照的亮度,从而增强图像㊂此类方法不仅提高了图像的对比度,还降低了噪声带来的的影响,不足之处是要根据经验来人工设置算法的相关参数,并且不能对不同类型图像自适应增强㊂基于深度学习的方法,Lore等[6]提出的LLNet通过类深度神经网络来识别低光图像中的信号特征并对其自适应增强;Wei等[7]提出的Retinex-Net结合Retinex理论和神经网络进行图像增强;Wang等[8]提出的GLADNet先通过编解码网络对低光图像生成全局照明先验知识,然后结合全局照明先验知识和输入图像,采用卷积神经网络来增强图像的细节㊂此类基于深度学习的方法均取得了不错的效果,但是大多数方法在增强过程中并没有对噪声进行有效抑制,从而使得增强后的图像出现噪声大㊁颜色失真等问题㊂为解决这些问题,提出了Retinex-RANet(Retinex-Residuals Attention Net)方法㊂Retinex-RANet首先在分解阶段采用残差模块与跳跃连接,通过跳跃连接将第一个卷积层提取的特征与每一个RB提取的特征融合,从而得到更准确的反分量和光照分量㊂另外,还在降噪网络中加入通道注意力模块和空洞卷积,注意力机制可以更好地去除反射分量中的噪声,还原细节;而空洞卷积能获取更多的上下文信息特征㊂实验结果表明:Retinex-RANet具有更好的低光图像增强效果㊂2㊀模㊀型Retinex-RANet模型框图如图1所示㊂由图1可以看到:整个网络模型由3个子网络组成,即分解网络㊁降噪网络以及调整网络,分别用于分解图像㊁降低噪声和调整亮度㊂具体地说,首先该算法通过分解网络将低光照图像S l和正常光图像S h分解为反射分量(R l㊁R h)和光照分量(I l㊁I h),然后降噪网络将分解的反射分量R l作为输入,并使用光照分量I l作为约束来抑制反射分量中的噪声,同时将光照分量I l送入调整网络,来提高光照分量的对比度,最后融合Rᶄ和Iᶄ得到增强后的图像㊂输入Sh分解网络反射分量Rh光照分量Ih反射分量Rl光照分量Il分解网络输入Sl降噪网络调整网络输出R′I′Conv+ReluConv+ReluConv+ReluConvSigmoid图1㊀Retinex-RANet模型结构图Fig.1㊀Retinex-RANet model structure diagram2.1㊀分解网络基于Retinex理论方法的关键是在分解阶段如何得到高质量的光照分量和反射分量,而分解后的结果对后续的增强和降噪操作都会产生影响,因此,设计一个有效的网络对低光图像进行分解是很有必要的㊂分解网络结构如图2所示㊂输入Sl33Conv+Relu3?3Conv+Relu3?3Conv+Relu3?3Conv+Relu3?3Conv+Relu3?3Conv3?3Conv+Relu3?3Conv+Relu3?3Conv+Relu1?1Conv+Relu1?1Conv+Relu跳跃连接S C反射分量Rl光照分量IlS i g m o i d R BR B R B R B图2㊀分解网络Fig.2㊀Decomposition network在分解网络结构中,为了使深度神经网络在训练26第6期史宇飞,等:基于Retinex 理论的低光图像增强算法阶段更容易优化,使用3个残差块(RB)来获得更好的分解结果㊂首先使用3ˑ3卷积提取输入低光图像S l 的特征;然后再经过3个RB 模块提取更多的纹理㊁细节等特征,同时为了减少底层颜色㊁边缘线条等特征的丢失,引入了跳跃连接,即将第一个卷积层的输出连接到每一个RB 的输出,保证特征的充分提取;最后通过3ˑ3ˑ4的卷积层和sigmoid 函数激活,从而得到3通道的反射分量和1通道的光照分量㊂2.2㊀降噪网络在对低光图像进行增强的过程中,大多数基于Retinex 理论的方法在得到分解结果后都忽略了噪声的影响,这会导致最终的增强结果受到反射分量中噪声的干扰,出现模糊㊁失真等问题㊂为了解决这类问题,设计了如图3所示的降噪网络㊂输入R l 、I lS ES ES ES E S E (D i l a t e dC o n v +R e l u )2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2m a x p o o l i n g(C o n v +R e l u )?2C o n vG l o b a lA v g .p o o lF CF CR e l uS i g m o i dS c a l eW ?H ?CW ?H ?C S i g m o i d输出R ′11C图3㊀降噪网络Fig.3㊀Denoising network㊀㊀在低光增强领域,U-Net 网络由于其优秀的结构设计,被大量网络作为其主要架构和部分架构,因此Retinex-RANet 也采用U-Net 作为降噪网络的基础网络部分㊂降噪网络包含编码和解码两个部分㊂在编码阶段,先融合输入的反射分量和光照分量,然后经过一组由两个3ˑ3的空洞卷积㊁RELU 函数激活和最大池化层组成的编码块,3组均由两个卷积核为3ˑ3的卷积激活层和一个最大池化层组成的编码块来提取特征,从而得到编码阶段的特征图,最后将其送入解码阶段㊂编码过程中,每次通过一个编码块,图像的通道数会翻倍,但是其尺寸会降低一半㊂在解码阶段,由4个相同的解码块组成,结构为3ˑ3的卷积层 RELU 函数激活 2ˑ2的反卷积层㊂受到图像识别中的SENet [9]的启发,将通道注意力模块嵌入到跳跃连接中,以便更好地降低噪声,恢复细节㊂如图3所示:首先将编码阶段采集到的图像特征进行全局平均池化操作,然后经过两个全连接层和两个激活函数,最后和解码阶段的特征图逐通道相乘,此过程可将更多的权重分配给有用的特征,如颜色㊁细节和纹理特征等,同时为噪声㊁阴影快和伪影等特征分配较少的权重;然后融合跳跃连接得到的特征图与反卷积后的特征,之后再进行卷积计算,解码过程中,每次通过一个解码块,图像的通道数会降低一半,但是其尺寸会翻倍;最后使用3ˑ3卷积得到一个3通道特征图,并对其进行sigmoid 函数激活,从而得到降噪后的反射分量㊂2.3㊀调整网络在得到分解后的光照分量后,需要提高其对比度,因此设计了图1中的调整网络㊂此调整网络是一个轻量级网络,包含3个卷积激活层㊁1个卷积层和1个Sigmoid 层,同时为了避免底层信息的损失,通过跳跃连接将输入连接到最后一个卷积层的输出㊂2.4㊀损失函数训练时,3个子网络均单独训练,因此,整个Retinex-RANet 的损失由分解损失L dc ㊁降噪损失L r 和调整损失L i 组成㊂2.4.1㊀分解损失为了更好地从低光图像中分解出反射分量和光照分量,设计了3个损失函数,即重建损失L rec ㊁反射分量36重庆工商大学学报(自然科学版)第40卷一致性损失L rs ㊁光照分量平滑损失L is ,如下所示:L dc =L rec +λ1L rs +λ2L isL rec = S l -R l I l 1+ S h -R h I h 1L rs = R l -R h 22L is =∇I lmax ∇S l ,ε()1+∇I hmax ∇S h ,ε()1其中,λ1和λ2分别为L rs 和L is 的权重系数,S l 和S h 为低光条件和正常光条件下的输入图像,R l ㊁R h 和I l ㊁I h 分别是低光和正常光图像分解后的反射分量和光照分量, 表示逐像素相乘操作, 1表示使用的是L 1范数约束损失, 2表示使用的是L 2范数约束损失,∇表示梯度,为水平梯度与垂直梯度之和,ɛ为一个小的正常数,取0.01㊂2.4.2㊀降噪损失为了保证经过降噪处理后的反射分量和正常光图像的反射分量在结构㊁纹理信息等方面能够保持一致,同时衡量降噪处理后图像与正常光图像之间的颜色差异,降噪网络的损失函数L r 如下所示:L r = R ᶄ-R h 22-SSIM R ᶄ,R h ()+ ∇R ᶄ-∇R h 22+L cR ᶄ为经过降噪处理后的反射率,SSIM ()为结构相似性度量,L c 为色彩损失函数,表达式如下:L c = ΓR ᶄ()-ΓS h () 22其具体含义为先对降噪后的图像R ᶄ和正常光图像S h 进行高斯模糊Γ(),再计算模糊后图像的均方误差㊂2.4.3㊀调整损失为了使调整过后的光照分量与正常光图像的光照分量尽可能相似,调整网络的损失函数L i 如下所示:L i = I ᶄ-I h 22+ ∇I ᶄ-∇I h 22其中,I ᶄ为I l 增强后的图像㊂3㊀实验结果和分析3.1㊀训练数据集实验中的训练集为LOL 数据集[7],该数据集包含500对图像:其中,训练集含485对图像,验证集为剩余15对图像㊂在训练过程中,分解模块和增强模块的批量化大小为16,块大小为48ˑ48,训练次数为2000次,分解网络损失函数的权重系数λ1=0.01,λ2=0.2㊂降噪模块的批量化大小为4,块大小为384ˑ384,训练次数为1000次㊂模型优化方法为随机梯度下降法㊂整个网络模型在CPU 型号为Intel (R )Core (TM )i7-10700K㊁GPU 型号为Nvidia GeForce RTX 2080Ti 的电脑上运行,同时训练框架为Tensorflow1.15,GPU 使用Nvidia CUDA10.0和CuDNN7.6.5加速㊂为了评估Retinex-RANet 的性能,将其与几种传统方法如BIMEF [10]㊁Dong [11]㊁LIME [12]㊁MF [13]㊁MSR [14]和SIRE [15]等,以及深度学习方法,如R2RNet [16]㊁Retinex-Net [7]㊁KinD [17]㊁Zero-Dce [18]等进行比较,并同时在多个数据集上评估了该算法,包括LOL㊁LIME㊁NPE [19]和MEF [20]数据集㊂在实验过程中,均采用原文献所提供的源代码对图像进行训练和测试㊂在评估过程中,采用峰值信噪比(R PSNR [21])㊁结构相似性(R SSIM [22])和自然图像质量评估(R NIQE [23])这3个指标来进行定量比较㊂R PSNR 和R SSIM 值越高,R NIQE 值越低,则增强后图像的质量越好㊂3.2㊀消融实验为了确定Retinex-RANet 的有效性,在KinD 网络的基础上进行消融实验㊂该实验使用LOL 数据集进行验证,同时采用R PSNR ㊁R SSIM 指标来评估增强后图像的质量㊂结果如表1所示,表中RB 表示残差模块,SC 表示跳跃连接,SE 表示注意力模块㊂表1㊀各改进模块的消融实验结果Table 1㊀Ablation experimental results of each improved module序号算法R PSNRR SSIM1KinD16.12450.71132KinD +RB 17.27750.76633KinD +RB +SC17.76050.77934KinD +RB +SC +SE19.01960.78395Ours19.77610.7922表1中序号2给出的是在KinD 网络基础上,使用残差模块作为分解网络时的结果㊂相比于KinD 网络,R PSNR 和R SSIM 均有显著的提升㊂在此基础上加入跳跃连接,见序号3,相较于序号2的结果又有了小幅提升㊂说明在使用残差模块和跳跃连接作为分解网络的情况下,得到了质量更高的分解结果,从而验证了残差模块和跳跃连接的有效性㊂由于3个子网络是单独进行训练的,确定改进的分解网络有用后,在此基础上确定在降噪网络中加入空洞卷积和注意力机制的有用性㊂从序号4的结果可以看出,在加入注意力机制后,图像指标明显上升,这是因为注意力模块能集中学习有用特征,如颜色㊁细节等,从而降低图像中的噪声,阴影等㊂为了获取更多的上下文信息,同时在降噪网络中加入空洞卷积(序号5),相较于序号4的结果有了小幅提升㊂从而确定了Retinex-RANet 的模型即为序号5的模型㊂3.3㊀实验评估各算法在不同数据集上的视觉对比如图4㊁图5所示㊂46第6期史宇飞,等:基于Retinex 理论的低光图像增强算法I n p u t D o n g R e t i n e x N e t Z e r o D ceM S R L I M E S I R E R 2R N etM F K i n DO u r sG r o u n d T r u th图4㊀LOL 数据集上各算法的视觉效果Fig.4㊀Visual effects of each algorithm on the LOL datasetI n p u t D o n g R e t i n e x N e t Z e r o D ceM S R L I M E S I R E R 2R N etM FB I M E F K i n DO u rs图5㊀其他数据集上各算法的视觉效果Fig.5㊀Visual effects of each algorithm on other datasets㊀㊀图4的输入来自LOL 数据集,是非常低亮度的真实世界图像㊂可以看出:Dong㊁Retinex-Net㊁Zero-Dce㊁MSR㊁LIME 的增强结果中存在明显的噪声㊁色差等问题,特别是对Retinex-Net 来说,看起来不像真实世界的56重庆工商大学学报(自然科学版)第40卷图像;SIRE㊁MF和KinD对图像的增量程度有限,增强结果偏暗;R2RNet的增强结果在整体上偏白,存在饱和度过低等问题;相比之下,Retinex-RANet增强后的图片更接近于真实世界图像,有效抑制了噪声,同时能很好地还原图像原有的色彩㊂此外,还在其他数据集上对本模型进行了测试,如图5所示㊂从左上角的细节图像中可以看到:虽然大多数方法都能在一定程度上改变输入图像的亮度,但仍然存在着一些严重的视觉缺陷,比如Dong和Retinex-Net存在噪声和颜色失真问题;Zero-Dce㊁R2RNet和MSR增强后的图像整体偏白,无法看清左上角图像的背景;SIRE和KinD增强后的图像总体偏暗,无法观察脸部细节;Retinex-RANet㊁LIME㊁MF和BIMEF 能相对清晰地观察到脸部细节,但比较左下角图的可知,Retinex-RANet相较于其他算法,增强的亮度适中,轮廓细节更加清晰,色彩更为真实㊂表2显示了在LOL数据集上各算法的评估对比,其中,加黑数字为最优数值㊂LOL数据集中的图像为成对的低光/正常光图像,因此可使用R PSNR和R SSIM 来衡量算法的优越性,同时还引用了R NIQE指标㊂从表中可以看出:在R PSNR和R SSIM指标上,Retinex-RANet相较于其他算法都取得了最高的值,而在R NIQE 指标上,所取得的值略高于KIND和R2RNet算法得到的值㊂因为LIME㊁NPE和MEF数据集只包含低光图像,无对应的正常光图像,所以只使用R NIQE指标来比较各算法之间的差异㊂从表3可以看出:在LIME和NPE数据集上,Retinex-RANet取得了最优值,而在MEF数据集上,所取得的值略高于SRIE算法得到的值㊂表2㊀LOL数据集上各算法的结果对比Table2㊀Comparison of the results of each algorithm on the LOL dataset指标SRIE MSR LIME Dong MF Zero-Dce Retinex-Net KinD R2RNet Ours R PSNR13.348612.097914.758315.263915.667616.361516.731716.124518.934219.7761R SSIM0.39760.36370.33610.34470.36890.52470.43090.71130.75250.7982 R NIQE7.28698.11368.37768.31578.77717.93138.8788 4.6724 3.7657 4.7465表3㊀不同数据集上的R NIQE对比Table3㊀Comparison of R NIQE on different datasets算法LIME-data NPE-data MEF-data SRIE 3.8596 4.1803 3.4456MSR 3.7642 4.0614 3.5654 LIME 3.7862 4.4466 3.7962 Dong 4.0516 4.6952 4.2759MF 4.0673 4.3506 3.5995 Zero-Dce 4.3421 4.6511 3.5532 Retinex-Net 4.8077 4.5712 5.1747 KinD 4.1441 3.933 4.7805 R2RNet 5.2291 4.0191 5.1082Ours 3.4064 3.4984 3.4621综上所述,虽然Retinex-RANet并没有在上述数据集上都取得最好的结果,但仍有一定优势㊂同时,在客观评判指标R SSIM和R PSNR上均取得了最高值㊂因此, Retinex-RANet相较于其他算法,对低光照图像增强后的效果更优㊂4㊀结束语针对低光图像在视觉效果上存在亮度低㊁噪声大以及对比度弱等问题,设计了Retinex-RANet网络模型㊂此模型在分解网络中结合残差模块(RB)和跳跃连接,充分提取图像特征和细节信息;在降噪网络中嵌入空洞卷积和注意力机制,可以获取更多的上下文信息,降低图像中的噪声㊁阴影等;最后将降噪网络去噪后的反射分量和亮度调整网络增强后的光照分量融合,得到最终的增强结果㊂实验表明:与LIME㊁Zero-Dce和R2RNet相比,Retinex-RANet在客观指标R PSNR 和R SSIM上均取得了最高的数值,Retinex-RANet在增强图像的视觉对比上,不仅提高了图像的对比度㊁抑制了噪声,而且明显消除了退化问题,达到了更好的视觉效果㊂66第6期史宇飞,等:基于Retinex理论的低光图像增强算法参考文献References1 ㊀SUBRAMANI B VELUCHAMY M.Fuzzy gray level differencehistogram equalization for medical image enhancement J .Journal of Medical Systems 2020 44 6 103 110.2 ㊀张驰谭南林李响等.基于改进型Retinex算法的雾天图像增强技术J .北京航空航天大学学报2019 452309 316.ZHANG Chi TAN Nan-lin LI Xiang et al.Foggy sky image enhancement technology based on the improved Retinex algorithm J .Journal of Beijing University of Aeronautics and Astronautics 2019 45 2 309 316.3 ㊀DONG X WANG G PANG Y et al.Fast efficient algorithmfor enhancement of low lighting video C .Barcelona ICME 2011 1 6.4 ㊀LI L WANG R WANG W et al.A low-light image enhancementmethod for both denoising and contrast enlarging C .QC Canda ICIP 2015 3730 3734.5 ㊀PARK S YU S KIM M et al.Dual autoencoder network forRetinex based low light image enhancement J .IEEE Access 2018 6 22084 22093.6 ㊀LORE K G AKINTAYO A SARKAR S.LLNet A deepautoencoder approach to natural low-light image enhancement J .Pattern Recognition 2017 61 650 662.7 ㊀CHEN W WANG W J YANG W H et al.Deep Retinexdecomposition for low-light enhancement C//Proceedings of British Machine Vision Conference BMVC .2018 155 158.8 ㊀WANG W J CHEN W YANG W H et al.GLADNet Low-light enhancement network with global awareness C .Xi anChina FG 2018 751 755.9 ㊀HU J SHEN L SUN G.Squeeze-and-excitation networks C .Salt Lake City UT USA CVPR 2018 7132 7141.10 YING Z GE L GAO W.A bio-inspired multi-exposurefusion framework for low-light image enhancement EB/OL .https ///abs/1711.0059/.2017.11 DONG J PANG Y WEN J.Fast efficient algorithm forenhancement of low lighting video C .Barcelona ICME 2011 1 6.12 GUO X LI Y LING H.Lime Low-light image enhancement via illumination map estimation J .IEEE Trans Image Process 2017 26 2 982 993.13 FU X ZENG D YUE H et al.A fusion-based enhancing method for weakly illuminated images J .Signal Processing 2016 129 82 96.14 JOBSON D J RAHMAN Z WOODELL G A.A multiscale Retinex for bridging the gap between color images and the human observation of scenes J .IEEE Transactions on Image processing 1997 6 7 965 976.15 FU X ZENG D HUANG Y et al.A weighted variational model for simultaneous reflectance and illumination estimation C .Las Vegas NV USA CVPR 2016 2782 2790.16 HAI J XUAN Z YANG R et al.R2RNet Low-light image enhancement via real-low to real-normal network EB/OL . https //arxivorg/ahs/2016.14501.2021.17 ZHANG Y ZHANG J GUO X.Kindling the darkness A practical low-light image enhancer C //27th ACM Multimedia. France Nice 2019 1632 1640.18 GUO C LI C GUO J et al.Zero-reference deep curve estimation for low-light image enhancement C .Seattle WA USA CVPR 2020 1780 1789.19 WANG S ZHENG J HU H M et al.Naturalness preserved enhancement algorithm for non-uniform illumination images J . IEEE Transactions on Image Processing 2013 229 3538 3548.20 MA J ZENG K WANG Z.Perceptual quality assessment for multi-exposure image fusion J .IEEE Transactions on Image Processing 2015 24 11 3345 3356.21 HUYNH-THU Q GHANBARI M.Scope of validity of PSNR in image/video quality assessment J .Electronics Letters 2008 44 13 800 801.22 WANG Z BOVIK A C SHEIKH H R et al.Image quality assessment From error visibility to structural similarity J .IEEE Transactions on Image Processing 2004 13 4 600 612.23 MITTAL A SOUNDARARAJAN R BOVIK A C.Making a completely blind image quality analyzer J .IEEE Signal Processing Letters 2012 20 3 209 212.责任编辑:李翠薇76。
湖南省长沙市雅礼中学2024届高三下学期3月综合测试(一)英语试题
湖南省雅礼中学2024届高三综合自主测试(一)英语试卷第一部分听力(共两节,满分30分)做题时,先将答案标在试卷上。
录音内容结束后,你将有两分钟的时间将试卷上的答案转涂到答题卡上。
第一节(共5小题;每小题1.5分,满分7.5分)1.Why does the woman intend to go to Rome?A. To work.B. To study.C. To travel.2.What does the woman think of the trip?A. Worthless.B. Terrible.C. Great.3.In which city did the woman and John stay the longest?A. Vienna.B. Rome.C. Paris.4.What is “couscous”?A.A new hotel.B.A kind of food.C.A close relative.5.Where does the conversation most probably take place?A. In a park.B. In a zoo.C. In a pet store.第二节(共15小题;每小题1.5分,满分22.5分)听下面5段对话或独白。
每段对话或独白后有几个小题,从题中所给的A、B、C三个选项中选出最佳选项,并标在试卷的相应位置。
听每段对话或独白前,你将有时间阅读各个小题,每小题5秒钟;听完后,各小题将给出5秒钟的作答时间。
每段对话或独白读两遍。
听下面一段对话,回答6-7小题。
6.Where are the speakers?A. In a restaurant.B. In a bookstore.C. In a supermarket.7.What does the man have to do now?A. Sign his name.B. Wait for his turn.C. Call his friend.听下面一段对话,回答8-10小题。
江苏省南通市海安高级中学2023-2024学年高一下学期第一次月考英语试题
2023-2024学年度第二学期高一年级阶段检测(一)英语第一部分听力(共两节,满分30分)第一节(共5小题;每小题1.5分,满分7.5分)听下面5段对话。
每段对话后有一个小题,从题中所给的A、B、C三个选项中选出最佳选项,并标在试卷的相应位置。
听完每段对话后,你都有10秒钟的时间来回答有关小题和阅读下一小题。
每段对话仅读一遍。
1.How long does the museum open each day?A.Eight hours.B.Seven hours.C.Six hours.2.How does the woman feel now?A.Ill.B.Sad.C.Angry.3.Who is the woman?A.The man’s teacher.B.The man’s coach.C.The man’s mother.4.What will the woman do?A.Give Joan a call.B.Tell Joan about the meeting.C.Have lunch with the man. 5.Where did Jane move to?A.A place in the country.B.A place out of London.C.A place with a good view.第二节(共15小题;每小题1.5分,满分22.5分)听下面5段对话或独白。
每段对话或独白后有几个小题,从题中所给的A、B、C三个选项中选出最佳选项,并标在试卷的相应位置。
听每段对话或独白前,你将有时间阅读各个小题,每小题5秒钟:听完后,各小题将给出5秒钟的作答时间。
每段对话或独白读两遍。
听下面一段对话,回答以下小题。
6.What are the two speakers doing now?A.Having dinner.B.Seeing a film C.Cleaning the table. 7.What is the man complaining about?A.The woman is too busy with her work.B.The woman hardly ever sees a movie with him.C.The woman watches too much TV and ignores him.听下面一段对话,回答以下小题。
基于边缘检测的抗遮挡相关滤波跟踪算法
基于边缘检测的抗遮挡相关滤波跟踪算法唐艺北方工业大学 北京 100144摘要:无人机跟踪目标因其便利性得到越来越多的关注。
基于相关滤波算法利用边缘检测优化样本质量,并在边缘检测打分环节加入平滑约束项,增加了候选框包含目标的准确度,达到降低计算复杂度、提高跟踪鲁棒性的效果。
利用自适应多特征融合增强特征表达能力,提高目标跟踪精准度。
引入遮挡判断机制和自适应更新学习率,减少遮挡对滤波模板的影响,提高目标跟踪成功率。
通过在OTB-2015和UAV123数据集上的实验进行定性定量的评估,论证了所研究算法相较于其他跟踪算法具有一定的优越性。
关键词:无人机 目标追踪 相关滤波 多特征融合 边缘检测中图分类号:TN713;TP391.41;TG441.7文献标识码:A 文章编号:1672-3791(2024)05-0057-04 The Anti-Occlusion Correlation Filtering Tracking AlgorithmBased on Edge DetectionTANG YiNorth China University of Technology, Beijing, 100144 ChinaAbstract: For its convenience, tracking targets with unmanned aerial vehicles is getting more and more attention. Based on the correlation filtering algorithm, the quality of samples is optimized by edge detection, and smoothing constraints are added to the edge detection scoring link, which increases the accuracy of targets included in candi⁃date boxes, and achieves the effects of reducing computational complexity and improving tracking robustness. Adap⁃tive multi-feature fusion is used to enhance the feature expression capability, which improves the accuracy of target tracking. The occlusion detection mechanism and the adaptive updating learning rate are introduced to reduce the impact of occlusion on filtering templates, which improves the success rate of target tracking. Qualitative evaluation and quantitative evaluation are conducted through experiments on OTB-2015 and UAV123 datasets, which dem⁃onstrates the superiority of the studied algorithm over other tracking algorithms.Key Words: Unmanned aerial vehicle; Target tracking; Correlation filtering; Multi-feature fusion; Edge detection近年来,无人机成为热点话题,具有不同用途的无人机频繁出现在大众视野。
基于避障路径规划的无人直升机空地跟踪控制
第44卷 第1期2024 年2月辽宁石油化工大学学报JOURNAL OF LIAONING PETROCHEMICAL UNIVERSITYVol.44 No.1Feb. 2024引用格式:杨静雯,李涛,杨欣,等.基于避障路径规划的无人直升机空地跟踪控制[J].辽宁石油化工大学学报,2024,44(1): 71-79.YANG Jingwen,LI Tao,YANG Xin,et al.Collaborative Air⁃Ground Tracking Control of Unmanned Helicopter Based on Obstacle Avoidance Path Planning[J].Journal of Liaoning Petrochemical University,2024,44(1):71-79.基于避障路径规划的无人直升机空地跟踪控制杨静雯,李涛,杨欣,冀明飞(南京航空航天大学自动化学院,江苏南京 211106)摘要: 针对无人直升机(Unmanned Aerial Helicopter,UAH)在空地协同跟踪过程中的避障和控制问题,提出了新型路径避障规划和跟踪控制设计方法。
针对不确定性的线性UAH模型,通过对UAH警示范围内二维环境信息进行处理判断,借助摸墙算法(Wall⁃Following Algorithm) 提出合适的避障策略,计算避障路径的行进角度以及能够弥补绕行距离的跟踪速度;将所得避障方法拓展至三维环境中,根据水平和垂直方向上的障碍物信息确定UAH飞行角度,从而减小由避障环节所带来的绕行距离;在上述避障算法的基础上,引入人工神经网络(Approximate Nearest Neighbor,ANN)估计模型不确定项,进而结合前馈补偿与最优控制等技术建立了跟踪控制设计方案。
仿真结果表明,所提避障策略和控制算法有效。
关键词: 无人直升机; 空地跟踪; 避障路径规划; 人工神经网络中图分类号: TP13 文献标志码: A doi:10.12422/j.issn.1672⁃6952.2024.01.011Collaborative Air⁃Ground Tracking Control of Unmanned Helicopter Based onObstacle Avoidance Path PlanningYANG Jingwen,LI Tao,YANG Xin,JI Mingfei(College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing Jiangsu 211106,China)Abstract: The paper aims to study the problem of obstacle avoidance in air⁃ground cooperative tracking control for the unmanned aerial helicopter (UAH),in which a new approach of designing the path obstacle avoidance plan and controller design is proposed. Initially, as for the uncertain linear UAH,by processing and judging two⁃dimensional environmental information within the warning range for the UAH,an obstacle avoidance strategy is proposed with the help of wall⁃following algorithm,and the flight angle of obstacle avoidance path and the tracking speed that can make up for bypass distance are calculated.Secondly,the proposed obstacle avoidance method is extended to the three⁃dimensional case,and the flight angle of the UAH is determined based on the obstacle information in the horizontal and vertical directions,which can reduce the bypass distance caused by the obstacle avoidance link as possible.Thirdly,based on two derived obstacle avoidance algorithms above,the artificial neural network (ANN) is introduced to estimate model uncertainty,and then the tracking control design schemes are established by using feedforward compensation and optimal control technologies.some simulations demonstrate the effectiveness of the proposed obstacle avoidance strategy and control algorithm.Keywords: Unmanned aerial helicopter ; Clearing tracking; Obstacle avoidance path planning; Artificial neural networ无人直升机(Unmanned Aerial Helicopter,UAH)是一种能够利用机载航空电子系统,通过无线电远距离遥控或无人干预下自主完成控制任务的飞行器。
大型光学红外望远镜拼接非球面子镜反衍补偿检测光路设计
大型光学红外望远镜拼接非球面子镜反衍补偿检测光路设计王丰璞 李新南 徐晨 黄亚Optical testing path design for LOT aspheric segmented mirrors with reflective-diffractive compensationWANG Feng-pu, LI Xin-nan, XU Chen, HUANG Ya引用本文:王丰璞,李新南,徐晨,黄亚. 大型光学红外望远镜拼接非球面子镜反衍补偿检测光路设计[J]. 中国光学, 2021, 14(5): 1184-1193. doi: 10.37188/CO.2020-0218WANG Feng-pu, LI Xin-nan, XU Chen, HUANG Ya. Optical testing path design for LOT aspheric segmented mirrors with reflective-diffractive compensation[J]. Chinese Optics, 2021, 14(5): 1184-1193. doi: 10.37188/CO.2020-0218在线阅读 View online: https:///10.37188/CO.2020-0218您可能感兴趣的其他文章Articles you may be interested in非零位凸非球面子孔径拼接检测技术研究Research on non-null convex aspherical sub-aperture stitching detection technology中国光学. 2018, 11(5): 798 https:///10.3788/CO.20181105.0798易测量非球面定义及应用Definition and application of easily measurable aspheric surfaces中国光学. 2017, 10(2): 256 https:///10.3788/CO.20171002.0256大偏离度非球面检测畸变校正方法Distortion correcting method when testing large-departure asphere中国光学. 2017, 10(3): 383 https:///10.3788/CO.20171003.0383基于单次傅里叶变换的分段衍射算法Step diffraction algorithm based on single fast Fourier transform algorithm中国光学. 2018, 11(4): 568 https:///10.3788/CO.20181104.0568一种针对超大口径凸非球面的面形检测方法Surface testing method for ultra-large convex aspheric surfaces中国光学. 2019, 12(5): 1147 https:///10.3788/CO.20191205.1147超颖表面原理与研究进展The principle and research progress of metasurfaces中国光学. 2017, 10(5): 523 https:///10.3788/CO.20171005.0523文章编号 2095-1531(2021)05-1184-10大型光学红外望远镜拼接非球面子镜反衍补偿检测光路设计王丰璞1,2,3,李新南1,2 *,徐 晨1,2,黄 亚1,2(1. 中国科学院 国家天文台 南京天文光学技术研究所,江苏南京 210042;2. 中国科学院天文光学技术重点实验室 (南京天文光学技术研究所),江苏南京 210042;3. 中国科学院大学,北京 100049)摘要:为了实现大口径、长焦距、批量化离轴镜面的高精度面形检验,本文提出了一种零位反衍补偿检测方案,采用计算全息和球面反射镜共同对离轴镜面法向像差进行补偿,检测光路波像差残差接近于零。
基于GCN-LSTM_的频谱预测算法
doi:10.3969/j.issn.1003-3114.2023.02.001引用格式:薛文举,付宁,高玉龙.基于GCN-LSTM 的频谱预测算法[J].无线电通信技术,2023,49(2):203-208.[XUE Wenju,FU Ning,GAO Yulong.Spectrum Prediction Algorithm Based on GCN-LSTM[J].Radio Communications Technology,2023,49(2):203-208.]基于GCN-LSTM 的频谱预测算法薛文举,付㊀宁,高玉龙(哈尔滨工业大学通信技术研究所,黑龙江哈尔滨150001)摘㊀要:无线频谱是一项重要的㊁难以再生的自然资源㊂在频谱数据中随着信道的动态变化,各个信道不能建模成规则的结构㊂由于卷积神经网络提取的是规则数据结构的相关性,没有考虑信道动态变化以及各个信道节点之间的相关性影响,基于此研究了基于图卷积神经网络(Graph Convolutional Network,GCN)和长短期记忆(Long Short-TermMemory,LSTM)网络结合的GCN-LSTM 频谱预测模型,并且引入了注意力机制,仿真得到了GCN-LSTM 在正确数据集和有一定错误数据的数据集上的预测性能和算法运行时间㊂结果表明在引入注意力机制后,GCN-LSTM 预测模型的准确性和实时性都得到了提高㊂关键词:频谱预测;图神经网络;LSTM;注意力机制中图分类号:TN919.23㊀㊀㊀文献标志码:A㊀㊀㊀开放科学(资源服务)标识码(OSID):文章编号:1003-3114(2023)02-0203-06Spectrum Prediction Algorithm Based on GCN-LSTMXUE Wenju,FU Ning,GAO Yulong(Communication Research Center,Harbin Institute of Technology,Harbin 150001,China)Abstract :Wireless spectrum is an important and hard-to-regenerate natural resource.Since convolutional neural network extractscorrelation of regular data structure,dynamic changes of channel and the correlation between each channel node are not considered.Therefore,this paper studies a GCN-LSTM spectrum prediction model based on the combination of graph convolution neural network GCN and LSTM network,and introduces an attention mechanism.Simulation results show that the prediction performance and algorithm running time of GCN-LSTM on the correct dataset and the dataset with certain error data.Results show that the accuracy and real-timeperformance of GCN-LSTM prediction model are improved after introducing the attention mechanism.Keywords :spectrum prediction;graph neural network;LSTM;attention mechanism收稿日期:2022-12-29基金项目:国家自然科学基金(62171163)Foundation Item :National Natural Science Foundation of China(62171163)0 引言随着无线通信事业的蓬勃发展,各种接入无线网的智能设备数量迅速增长[1],频谱资源趋于紧缺㊂传统的静态频谱分配方式不适配于需求日渐多样化的频谱环境,出现了大量的 频谱空洞 ,造成了频谱资源浪费㊂为解决频谱利用不足的问题,Mitola 在1999年提出了认知无线电(Cognitive Radio,CR)的概念[2]㊂频谱预测的核心就是挖掘并利用历史频谱数据的相关性特征㊂频谱预测可以分为预测信道的占用情况或者是预测用户的位置和传输功率两大类㊂本文主要针对第一类,即预测信道的占用情况㊂早期研究主要采用例如自回归模型[3]㊁隐马尔可夫模型[4]㊁模式挖掘等传统方法㊂随着神经网络的发展,人们开始将神经网络,比如循环神经网络(Recurrent Neural Network,RNN)[5]和长短期记忆网络(Long Short-Term Memory,LSTM)[6]用于预测,LSTM 网络有效缓解了梯度消失和梯度爆炸现象㊂此外,有很多学者对时频联合域频谱预测展开了研究㊂文献[7]利用频谱的这种相关性提出一种二维频繁模式挖掘算法㊂由于不同地点频谱的使用情况也会有很大不同,因此也有研究将频谱预测的维度扩展到时频空域上㊂文献[8]利用神经网络来进行多维频谱预测的方法研究,提出了LSTM网络和其他神经网络结合的方法进行时频空三维的预测,然而只是提出了想法,并没有实现,算法仍处于仿真阶段㊂图神经网络最早由Gori等人[9]提出㊂GCN广泛用于提取图结构的特征信息,从理论上可以将GCN分为基于谱域和空域两类㊂Bruna等人在2014年提出了第一代GCN[10],定义了图上的卷积方法图结构㊂基于空域的图卷积则没有借助谱图理论,可以直接在空域上操作,非常灵活㊂Petar等人在2018年提出了图注意力网络(Graph Attention Network, GAT)[11],在图卷积网络中使用注意力机制,为图结构中不同的节点赋以不同的权重也就是注意力系数,解决了图卷积神经网络(Graph Convolutional Network,GCN)必须提前知道完整图结构的不足㊂把数据处理成图结构之后,利用图神经网络来学习图结构形式的数据可以更有效地挖掘发现其内部特征和模式,与频谱预测的核心不谋而合,因此可以使用图神经网络来进行频谱预测㊂本文首先分析了频谱预测的特点和发展趋势,说明了频谱预测的重要性和可行性㊂其次,针对频谱预测问题提出了GCN-LSTM模型进行二维时频频谱的预测,采用GCN提取频谱数据的拓扑特征,提取得到频谱数据中的频率相关性之后㊂然后利用LSTM网络进行时间维度动态性特征的提取㊂最后,通过引入注意力机制对GCN-LSTM频谱预测算法进行了改进研究㊂1 基于GCN-LSTM网络的频谱预测问题建模㊀㊀图神经网络可以通过分析研究各个节点的空间特征信息得到既包含内容也包含结构的特征表示,因此在本文中处理频谱数据时,不再是建模成规则的图片,而是建模成如图1所示的图结构㊂图结构中的每个节点代表频谱中的各个信道,信道之间是存在关联的,用图中的边表示,时间维度上的各个信道状态即是各个节点的特征㊂图1㊀频谱建模成图结构Fig.1㊀Spectrum modeling and mapping structure为了提取非欧式拓扑图的空间特征,研究人员利用GCN通过图结构的信息和图中节点的信息提取图的结构特征[12],如图2所示㊂GCN如今已经广泛应用于图数据的研究处理领域[13]㊂图2㊀图神经网络的结构示意图Fig.2㊀Structure diagram of graph neural network对于给定的图G=(V,E),V表示图中的节点集合,假设其长度为N㊂可以用图中的节点V和边E来对图进行定义㊂第二代图卷积GCN公式可以简化成:x G∗gθʈðK k=0θk T k(L~)x㊂(1)㊀㊀由式(1)可以看出,图上的卷积不需要整个图都参与运算,只需捕捉到图上的局部特征,减少了需要训练学习的参数量;并且不再需要对图进行特征分解,避免了特征分解的高昂代价㊂但是由于进行矩阵相乘操作,计算的时间复杂度仍然比较高㊂为了对问题进行简化,Kipf等人在文献[14]中设置K=1,只考虑节点的一阶邻居节点㊂如图3所示,当K=1时,对每个节点的特征进行更新时,不但会考虑各个节点本身的输入特征,还会将各个节点的一阶邻域的邻居节点的输入特征也考虑在内㊂取λmax =2,K =1,得到多层传播的图卷积计算公式:H (l +1)=σD ~-12A ~D ~-12H (l )W (l )(),(2)式中,σ(㊃)为非线性激活函数,A ~=A +I N ,A ~为加上自身属性后的邻接矩阵,D ~=ðjA ~ij 表示邻接矩阵A ~的度矩阵,H (l )为第l 层中图节点特征,H (0)=χ,即输入的特征矩阵,W (l )为第l 层的权重,即可训练的卷积滤波参数㊂图3㊀图卷积计算的简单示意图Fig.3㊀Simple diagram of convolution calculation2㊀增加注意力机制的GCN-LSTM 频谱预测算法2.1㊀GCN-LSTM 网络模型利用信道占用模型产生频谱数据,然后将频谱建模成图,频谱中的各个信道建模成图中的各个节点,在频率上提取信道之间的相关性即是提取节点之间的相关性,用GCN 进行提取,时间上的相关性则由LSTM 进行提取㊂GCN-LSTM 频谱预测算法示意如图4所示,内部结构如图5所示㊂图4㊀GCN-LSTM 模型示意图Fig.4㊀GCN-LSTM modeldiagram图5㊀GCN-LSTM 模型内部结构Fig.5㊀Internal structure diagram of GCN-LSTM model图4中,先将图结构形式的频谱输入GCN,提取其拓扑结构特征(即频率相关性),GCN 的输出Z N t 是已经提取了频率相关性的序列数据;然后将提取频率相关性的Z N t 序列输入进LSTM 网络,提取序列数据的时序相关性;最终通过激活函数的激活得到输出,并与真实的频谱数据利用损失函数衡量比较得到误差㊂Z N t 代表输入数据χt 经过图卷积网络后的数据特征㊂i t ㊁f t ㊁o t 分别代表了输入门(Input Gate)㊁遗忘门(Forget Gate)和输出门(Output Gate)㊂图5所示的χt 代表输入的处理成图结构的频谱数据,节点之间的关联强弱代表信道相关性的强弱㊂GCN-LSTM 预测模型公式如下:i t=σ(W iχ㊃Z Nt +W ih ㊃h t -1+b t )f t =σ(W f χ㊃Z N t +W fh ㊃h t -1+b f )o t =σ(W o χ㊃Z N t +W o h ㊃h t -1+b o )c ~t =g (W c χ㊃Z N t +W ch ㊃h t -1+b c )c t =i t☉c ~t +f t ☉c t -1h t =o t☉h -(c t )ìîíïïïïïïïï㊂(3)2.2㊀增加注意力机制的GCN-LSTM 预测模型注意力机制[15]是关注更重点的信息而忽略一些无关的信息,在GCN-LSTM 模型基础上,加入注意力机制,就是对不同时间步的特征赋予不同的权重㊂Soft Attention 注意力机制示意如图6所示,可以分成三步:一是信息输入h j ;二是注意力系数e ij 的计算,e ij 利用神经网络计算,再利用softmax 函数对e ij 进行归一化得到注意力的分布a ij ;三是利用注意力分布αij 与输入的信息进行加权平均得到输出c i㊂αij =exp(e ij )ðN k =1exp(eik)㊂(4)㊀㊀输出c i 为权重与输入的加权平均:c i =ðN j =1αijh j㊂(5)图6㊀Soft attention 注意力机制示意F i g.6㊀Schematic diagram of Soft Attention mechanism㊀㊀增加了注意力机制的GCN-LSTM 模型网络,如图7所示㊂将GCN-LSTM 的输出作为注意力层的输入,通过一个全连接层,再经过softmax 归一化,计算对时间步的权重即注意力分配矩阵,将注意力分配矩阵和输入数据进行逐元素的相乘即得到注意力的输出㊂图7㊀增加注意力机制的GCN-LSTM 模型示意图Fig.7㊀Schematic diagram of GCN-LSTM model forincreasing attention mechanism3 仿真结果利用信道占用模型,产生了5个信道的频谱数据,时间长度为10000,损失函数选择二分类交叉熵损失函数㊂在实验中,设置GCN 的模型参数为:图卷积网络层数为1,初始学习率为0.001,评价GCN-LSTM 预测算法的性能指标为准确率㊂预测窗口长度为10,隐藏单元数hidden_units 为128,batch_size 为64,迭代次数epoch 为20㊂基于GCN-LSTM 预测算法预测的准确率如图8和图9所示㊂图8㊀GCN-LSTM 模型准确率Fig.8㊀GCN-LSTM modelaccuracy图9㊀增加注意力机制的GCN-LSTM 模型精确率Fig.9㊀Increase the accuracy of GCN-LSTMmodel of attention mechanism二分类交叉熵binary_cross entropy 公式为:loss (y ,y ^)=-1nðni(y i lb(y^i )+(1-y i )lb(1-y ^i )),(6)式中,y i 为真实的值,y^i 为预测的值㊂在基础的GCN-LSTM 模型上增加了注意力机制之后,同样训练20轮之后,准确率从96.89%增长到97.86%,准确率得到了提升,训练时间从10.23s 变为12.69s,网络输出时间从0.13s 变为0.15s,时间基本为原来的1.19倍㊂这是因为增加注意力机制后,训练的参数数量从70020增长为78120,数量增多㊂增加注意力机制确实可以提高GCN-LSTM 模型整体的预测性能,而且性能略平稳一些㊂同时对比在频谱数据出现错误情况下的GCN-LSTM 和增加了注意力机制之后的预测模型的预测性能㊂图10为错误概率为0.05的情况,图11为错误概率为0.1的情况㊂比较无错误㊁错误概率为0.05和0.1时,随着错误概率的增加,准确率会略有下降㊂增加注意力机制后的预测算法比没有增加注意力机制的GCN-LSTM 算法指标提高一点,预测性能更好㊂图10㊀GCN-LSTM 模型错误率为0.05时的准确率Fig.10㊀Accuracy when GCN-LSTM model error rate is 0.05图11㊀GCN-LSTM 模型错误率为0.1时的准确率Fig.11㊀Accuracy when GCN-LSTM model error rate is 0.14 结论本文主要研究了基于GCN-LSTM 的频谱预测算法,采用GCN 和LSTM 复合网络GCN-LSTM 预测模型进行时频频谱预测㊂为了考量不同时间步的重要程度,在GCN-LSTM 预测模型基础上增加了注意力机制来提高预测效果㊂此外,实际数据可能存在错误的情况,对无错误数据和错误数据的情况分别进行了仿真㊂仿真结果表明,GCN-LSTM 方法预测准确率较高,且训练时间和预测时间更短,实时性大大提升㊂另外,增加注意力机制后,预测性能也得到一些提高,时间约是没增加注意力机制时的1.2倍㊂对比数据出现错误的情况下,使用GCN-LSTM 算法的预测性能也在可以接受的范围内㊂参考文献[1]㊀DEHOS C,GONZÁLEZ J L,DOMENICO A D,et -limeter-wave Access Andbackhauling:The Solution to the Exponential Data Traffic Increase in 5G Mobilecommuni-cations Systems [J ].IEEE Communications Magazine,2014,52(9):88-95.[2]㊀MITOLA J,MAGUIRE G Q.Cognitive Radio:MakingSoftware Radios More Personal[J].IEEE Personal Com-munications,1999,6(4):13-18.[3]㊀WEN Z,LUO T,XIANG W,et al.Autoregressive Spec-trum Hole Prediction Model for Cognitive Radio Systems [C]ʊIEEE International Conference on Communications Workshops.Beijing:IEEE,2008:154-157.[4]㊀何竞帆.认知无线电频谱预测算法研究[D].成都:电子科技大学,2019.[5]㊀邢玲.基于递归神经网络的频谱预测技术研究[D].成都:电子科技大学,2019.[6]㊀YU L,CHEN J,DING G.Spectrum Prediction via LongShort Term Memory [C]ʊ20173rd IEEE InternationalConference on Computer and Communications (ICCC).Chengdu:IEEE,2017:643-647.[7]㊀YIN S,CHEN D,ZHANG Q,et al.Mining SpectrumUsage Data:A Large-scale Spectrum Measurement Study[J].IEEE Transactions on Mobile Computing,2012,11(6):1033-1046.[8]㊀周佳宇,吴皓.基于神经网络的多维频谱推理方法探讨[J].移动通信,2018,42(2):35-39.[9]㊀GORI M,MONFARDINI G,SCARSELLI F.A New Modelfor Learning in Graph Domains [C]ʊProceedings 2005IEEE International Joint Conference on Neural Networks.Montreal:IEEE,2005:729-734.[10]BRUNA J,ZAREMBA W,SZLAM A,et al.Spectral Net-works and Deep Locally Connected Networks on Graphs [J /OL].arXiv:1312.6203[2022-12-20].https:ʊ /abs /1312.6203.[11]VELIC㊅KOVIC'P,CUCURULL G,CASANOVA A,et al.Graph Attention Networks[J/OL].arXiv:1710.10903[2022-12-20].https:ʊ/abs/1710.10903.[12]魏金泽.基于时空图网络的交通流预测方法研究[D].大连:大连理工大学,2021.[13]SCHLICHTKRULL M,KIPF T N,BLOEM P,et al.Mod-eling Relational Data with Graph Convolutional Networks[C]ʊEuropean Semantic Web Conference.Heraklion:Springer,2018:593-607.[14]KIPF T N,WELLING M.Semi-supervised Classificationwith Graph Convolutional Networks[J/OL].arXiv:1609.02907[2022-12-20].https:ʊ/abs/1609.02907.[15]UNGERLEIDER L G,KASTNER S.Mechanisms of VisualAttention in the Human Cortex[J].Annual Review ofNeuroscience,2003,23(1):315-341.作者简介:㊀㊀薛文举㊀哈尔滨工业大学硕士研究生㊂主要研究方向:频谱预测㊂㊀㊀付㊀宁㊀哈尔滨工业大学硕士研究生㊂主要研究方向:频谱预测㊂㊀㊀高玉龙㊀哈尔滨工业大学教授,博士生导师㊂主要研究方向:智能通信㊁频谱态势认知㊁智能信息融合㊂。
Advances in
Advances in Geosciences,4,17–22,2005 SRef-ID:1680-7359/adgeo/2005-4-17 European Geosciences Union©2005Author(s).This work is licensed under a Creative CommonsLicense.Advances in GeosciencesIncorporating level set methods in Geographical Information Systems(GIS)for land-surface process modelingD.PullarGeography Planning and Architecture,The University of Queensland,Brisbane QLD4072,Australia Received:1August2004–Revised:1November2004–Accepted:15November2004–Published:9August2005nd-surface processes include a broad class of models that operate at a landscape scale.Current modelling approaches tend to be specialised towards one type of pro-cess,yet it is the interaction of processes that is increasing seen as important to obtain a more integrated approach to land management.This paper presents a technique and a tool that may be applied generically to landscape processes. The technique tracks moving interfaces across landscapes for processes such as waterflow,biochemical diffusion,and plant dispersal.Its theoretical development applies a La-grangian approach to motion over a Eulerian grid space by tracking quantities across a landscape as an evolving front. An algorithm for this technique,called level set method,is implemented in a geographical information system(GIS).It fits with afield data model in GIS and is implemented as operators in map algebra.The paper describes an implemen-tation of the level set methods in a map algebra program-ming language,called MapScript,and gives example pro-gram scripts for applications in ecology and hydrology.1IntroductionOver the past decade there has been an explosion in the ap-plication of models to solve environmental issues.Many of these models are specific to one physical process and of-ten require expert knowledge to use.Increasingly generic modeling frameworks are being sought to provide analyti-cal tools to examine and resolve complex environmental and natural resource problems.These systems consider a vari-ety of land condition characteristics,interactions and driv-ing physical processes.Variables accounted for include cli-mate,topography,soils,geology,land cover,vegetation and hydro-geography(Moore et al.,1993).Physical interactions include processes for climatology,hydrology,topographic landsurface/sub-surfacefluxes and biological/ecological sys-Correspondence to:D.Pullar(d.pullar@.au)tems(Sklar and Costanza,1991).Progress has been made in linking model-specific systems with tools used by environ-mental managers,for instance geographical information sys-tems(GIS).While this approach,commonly referred to as loose coupling,provides a practical solution it still does not improve the scientific foundation of these models nor their integration with other models and related systems,such as decision support systems(Argent,2003).The alternative ap-proach is for tightly coupled systems which build functional-ity into a system or interface to domain libraries from which a user may build custom solutions using a macro language or program scripts.The approach supports integrated models through interface specifications which articulate the funda-mental assumptions and simplifications within these models. The problem is that there are no environmental modelling systems which are widely used by engineers and scientists that offer this level of interoperability,and the more com-monly used GIS systems do not currently support space and time representations and operations suitable for modelling environmental processes(Burrough,1998)(Sui and Magio, 1999).Providing a generic environmental modeling framework for practical environmental issues is challenging.It does not exist now despite an overwhelming demand because there are deep technical challenges to build integrated modeling frameworks in a scientifically rigorous manner.It is this chal-lenge this research addresses.1.1Background for ApproachThe paper describes a generic environmental modeling lan-guage integrated with a Geographical Information System (GIS)which supports spatial-temporal operators to model physical interactions occurring in two ways.The trivial case where interactions are isolated to a location,and the more common and complex case where interactions propa-gate spatially across landscape surfaces.The programming language has a strong theoretical and algorithmic basis.The-oretically,it assumes a Eulerian representation of state space,Fig.1.Shows a)a propagating interface parameterised by differ-ential equations,b)interface fronts have variable intensity and may expand or contract based onfield gradients and driving process. but propagates quantities across landscapes using Lagrangian equations of motion.In physics,a Lagrangian view focuses on how a quantity(water volume or particle)moves through space,whereas an Eulerian view focuses on a localfixed area of space and accounts for quantities moving through it.The benefit of this approach is that an Eulerian perspective is em-inently suited to representing the variation of environmen-tal phenomena across space,but it is difficult to conceptu-alise solutions for the equations of motion and has compu-tational drawbacks(Press et al.,1992).On the other hand, the Lagrangian view is often not favoured because it requires a global solution that makes it difficult to account for local variations,but has the advantage of solving equations of mo-tion in an intuitive and numerically direct way.The research will address this dilemma by adopting a novel approach from the image processing discipline that uses a Lagrangian ap-proach over an Eulerian grid.The approach,called level set methods,provides an efficient algorithm for modeling a natural advancing front in a host of settings(Sethian,1999). The reason the method works well over other approaches is that the advancing front is described by equations of motion (Lagrangian view),but computationally the front propagates over a vectorfield(Eulerian view).Hence,we have a very generic way to describe the motion of quantities,but can ex-plicitly solve their advancing properties locally as propagat-ing zones.The research work will adapt this technique for modeling the motion of environmental variables across time and space.Specifically,it will add new data models and op-erators to a geographical information system(GIS)for envi-ronmental modeling.This is considered to be a significant research imperative in spatial information science and tech-nology(Goodchild,2001).The main focus of this paper is to evaluate if the level set method(Sethian,1999)can:–provide a theoretically and empirically supportable methodology for modeling a range of integral landscape processes,–provide an algorithmic solution that is not sensitive to process timing,is computationally stable and efficient as compared to conventional explicit solutions to diffu-sive processes models,–be developed as part of a generic modelling language in GIS to express integrated models for natural resource and environmental problems?The outline for the paper is as follow.The next section will describe the theory for spatial-temporal processing us-ing level sets.Section3describes how this is implemented in a map algebra programming language.Two application examples are given–an ecological and a hydrological ex-ample–to demonstrate the use of operators for computing reactive-diffusive interactions in landscapes.Section4sum-marises the contribution of this research.2Theory2.1IntroductionLevel set methods(Sethian,1999)have been applied in a large collection of applications including,physics,chemistry,fluid dynamics,combustion,material science,fabrication of microelectronics,and computer vision.Level set methods compute an advancing interface using an Eulerian grid and the Lagrangian equations of motion.They are similar to cost distance modeling used in GIS(Burroughs and McDonnell, 1998)in that they compute the spread of a variable across space,but the motion is based upon partial differential equa-tions related to the physical process.The advancement of the interface is computed through time along a spatial gradient, and it may expand or contract in its extent.See Fig.1.2.2TheoryThe advantage of the level set method is that it models mo-tion along a state-space gradient.Level set methods start with the equation of motion,i.e.an advancing front with velocity F is characterised by an arrival surface T(x,y).Note that F is a velocityfield in a spatial sense.If F was constant this would result in an expanding series of circular fronts,but for different values in a velocityfield the front will have a more contorted appearance as shown in Fig.1b.The motion of thisinterface is always normal to the interface boundary,and its progress is regulated by several factors:F=f(L,G,I)(1)where L=local properties that determine the shape of advanc-ing front,G=global properties related to governing forces for its motion,I=independent properties that regulate and influ-ence the motion.If the advancing front is modeled strictly in terms of the movement of entity particles,then a straightfor-ward velocity equation describes its motion:|∇T|F=1given T0=0(2) where the arrival function T(x,y)is a travel cost surface,and T0is the initial position of the interface.Instead we use level sets to describe the interface as a complex function.The level set functionφis an evolving front consistent with the under-lying viscosity solution defined by partial differential equa-tions.This is expressed by the equation:φt+F|∇φ|=0givenφ(x,y,t=0)(3)whereφt is a complex interface function over time period 0..n,i.e.φ(x,y,t)=t0..tn,∇φis the spatial and temporal derivatives for viscosity equations.The Eulerian view over a spatial domain imposes a discretisation of space,i.e.the raster grid,which records changes in value z.Hence,the level set function becomesφ(x,y,z,t)to describe an evolv-ing surface over time.Further details are given in Sethian (1999)along with efficient algorithms.The next section de-scribes the integration of the level set methods with GIS.3Map algebra modelling3.1Map algebraSpatial models are written in a map algebra programming language.Map algebra is a function-oriented language that operates on four implicit spatial data types:point,neighbour-hood,zonal and whole landscape surfaces.Surfaces are typ-ically represented as a discrete raster where a point is a cell, a neighbourhood is a kernel centred on a cell,and zones are groups of mon examples of raster data include ter-rain models,categorical land cover maps,and scalar temper-ature surfaces.Map algebra is used to program many types of landscape models ranging from land suitability models to mineral exploration in the geosciences(Burrough and Mc-Donnell,1998;Bonham-Carter,1994).The syntax for map algebra follows a mathematical style with statements expressed as equations.These equations use operators to manipulate spatial data types for point and neighbourhoods.Expressions that manipulate a raster sur-face may use a global operation or alternatively iterate over the cells in a raster.For instance the GRID map algebra (Gao et al.,1993)defines an iteration construct,called do-cell,to apply equations on a cell-by-cell basis.This is triv-ially performed on columns and rows in a clockwork manner. However,for environmental phenomena there aresituations Fig.2.Spatial processing orders for raster.where the order of computations has a special significance. For instance,processes that involve spreading or transport acting along environmental gradients within the landscape. Therefore special control needs to be exercised on the order of execution.Burrough(1998)describes two extra control mechanisms for diffusion and directed topology.Figure2 shows the three principle types of processing orders,and they are:–row scan order governed by the clockwork lattice struc-ture,–spread order governed by the spreading or scattering ofa material from a more concentrated region,–flow order governed by advection which is the transport of a material due to velocity.Our implementation of map algebra,called MapScript (Pullar,2001),includes a special iteration construct that sup-ports these processing orders.MapScript is a lightweight lan-guage for processing raster-based GIS data using map alge-bra.The language parser and engine are built as a software component to interoperate with the IDRISI GIS(Eastman, 1997).MapScript is built in C++with a class hierarchy based upon a value type.Variants for value types include numeri-cal,boolean,template,cells,or a grid.MapScript supports combinations of these data types within equations with basic arithmetic and relational comparison operators.Algebra op-erations on templates typically result in an aggregate value assigned to a cell(Pullar,2001);this is similar to the con-volution integral in image algebras(Ritter et al.,1990).The language supports iteration to execute a block of statements in three ways:a)docell construct to process raster in a row scan order,b)dospread construct to process raster in a spreadwhile(time<100)dospreadpop=pop+(diffuse(kernel*pop))pop=pop+(r*pop*dt*(1-(pop/K)) enddoendwhere the diffusive constant is stored in thekernel:Fig.3.Map algebra script and convolution kernel for population dispersion.The variable pop is a raster,r,K and D are constants, dt is the model time step,and the kernel is a3×3template.It is assumed a time step is defined and the script is run in a simulation. Thefirst line contained in the nested cell processing construct(i.e. dospread)is the diffusive term and the second line is the population growth term.order,c)doflow to process raster byflow order.Examples are given in subsequent sections.Process models will also involve a timing loop which may be handled as a general while(<condition>)..end construct in MapScript where the condition expression includes a system time variable.This time variable is used in a specific fashion along with a system time step by certain operators,namely diffuse()andfluxflow() described in the next section,to model diffusion and advec-tion as a time evolving front.The evolving front represents quantities such as vegetation growth or surface runoff.3.2Ecological exampleThis section presents an ecological example based upon plant dispersal in a landscape.The population of a species follows a controlled growth rate and at the same time spreads across landscapes.The theory of the rate of spread of an organism is given in Tilman and Kareiva(1997).The area occupied by a species grows log-linear with time.This may be modelled by coupling a spatial diffusion term with an exponential pop-ulation growth term;the combination produces the familiar reaction-diffusion model.A simple growth population model is used where the reac-tion term considers one population controlled by births and mortalities is:dN dt =r·N1−NK(4)where N is the size of the population,r is the rate of change of population given in terms of the difference between birth and mortality rates,and K is the carrying capacity.Further dis-cussion of population models can be found in Jrgensen and Bendoricchio(2001).The diffusive term spreads a quantity through space at a specified rate:dudt=Dd2udx2(5) where u is the quantity which in our case is population size, and D is the diffusive coefficient.The model is operated as a coupled computation.Over a discretized space,or raster,the diffusive term is estimated using a numerical scheme(Press et al.,1992).The distance over which diffusion takes place in time step dt is minimally constrained by the raster resolution.For a stable computa-tional process the following condition must be satisfied:2Ddtdx2≤1(6) This basically states that to account for the diffusive pro-cess,the term2D·dx is less than the velocity of the advancing front.This would not be difficult to compute if D is constant, but is problematic if D is variable with respect to landscape conditions.This problem may be overcome by progressing along a diffusive front over the discrete raster based upon distance rather than being constrained by the cell resolution.The pro-cessing and diffusive operator is implemented in a map al-gebra programming language.The code fragment in Fig.3 shows a map algebra script for a single time step for the cou-pled reactive-diffusion model for population growth.The operator of interest in the script shown in Fig.3is the diffuse operator.It is assumed that the script is run with a given time step.The operator uses a system time step which is computed to balance the effect of process errors with effi-cient computation.With knowledge of the time step the it-erative construct applies an appropriate distance propagation such that the condition in Eq.(3)is not violated.The level set algorithm(Sethian,1999)is used to do this in a stable and accurate way.As a diffusive front propagates through the raster,a cost distance kernel assigns the proper time to each raster cell.The time assigned to the cell corresponds to the minimal cost it takes to reach that cell.Hence cell pro-cessing is controlled by propagating the kernel outward at a speed adaptive to the local context rather than meeting an arbitrary global constraint.3.3Hydrological exampleThis section presents a hydrological example based upon sur-face dispersal of excess rainfall across the terrain.The move-ment of water is described by the continuity equation:∂h∂t=e t−∇·q t(7) where h is the water depth(m),e t is the rainfall excess(m/s), q is the discharge(m/hr)at time t.Discharge is assumed to have steady uniformflow conditions,and is determined by Manning’s equation:q t=v t h t=1nh5/3ts1/2(8)putation of current cell(x+ x,t,t+ ).where q t is theflow velocity(m/s),h t is water depth,and s is the surface slope(m/m).An explicit method of calcula-tion is used to compute velocity and depth over raster cells, and equations are solved at each time step.A conservative form of afinite difference method solves for q t in Eq.(5). To simplify discussions we describe quasi-one-dimensional equations for theflow problem.The actual numerical com-putations are normally performed on an Eulerian grid(Julien et al.,1995).Finite-element approximations are made to solve the above partial differential equations for the one-dimensional case offlow along a strip of unit width.This leads to a cou-pled model with one term to maintain the continuity offlow and another term to compute theflow.In addition,all calcu-lations must progress from an uphill cell to the down slope cell.This is implemented in map algebra by a iteration con-struct,called doflow,which processes a raster byflow order. Flow distance is measured in cell size x per unit length. One strip is processed during a time interval t(Fig.4).The conservative solution for the continuity term using afirst or-der approximation for Eq.(5)is derived as:h x+ x,t+ t=h x+ x,t−q x+ x,t−q x,txt(9)where the inflow q x,t and outflow q x+x,t are calculated in the second term using Equation6as:q x,t=v x,t·h t(10) The calculations approximate discharge from previous time interval.Discharge is dynamically determined within the continuity equation by water depth.The rate of change in state variables for Equation6needs to satisfy a stability condition where v· t/ x≤1to maintain numerical stabil-ity.The physical interpretation of this is that afinite volume of water wouldflow across and out of a cell within the time step t.Typically the cell resolution isfixed for the raster, and adjusting the time step requires restarting the simulation while(time<120)doflow(dem)fvel=1/n*pow(depth,m)*sqrt(grade)depth=depth+(depth*fluxflow(fvel)) enddoendFig.5.Map algebra script for excess rainfallflow computed over a 120minute event.The variables depth and grade are rasters,fvel is theflow velocity,n and m are constants in Manning’s equation.It is assumed a time step is defined and the script is run in a simulation. Thefirst line in the nested cell processing(i.e.doflow)computes theflow velocity and the second line computes the change in depth from the previous value plus any net change(inflow–outflow)due to velocityflux across the cell.cycle.Flow velocities change dramatically over the course of a storm event,and it is problematic to set an appropriate time step which is efficient and yields a stable result.The hydrological model has been implemented in a map algebra programming language Pullar(2003).To overcome the problem mentioned above we have added high level oper-ators to compute theflow as an advancing front over a land-scape.The time step advances this front adaptively across the landscape based upon theflow velocity.The level set algorithm(Sethian,1999)is used to do this in a stable and accurate way.The map algebra script is given in Fig.5.The important operator is thefluxflow operator.It computes the advancing front for waterflow across a DEM by hydrologi-cal principles,and computes the local drainageflux rate for each cell.Theflux rate is used to compute the net change in a cell in terms offlow depth over an adaptive time step.4ConclusionsThe paper has described an approach to extend the function-ality of tightly coupled environmental models in GIS(Ar-gent,2004).A long standing criticism of GIS has been its in-ability to handle dynamic spatial models.Other researchers have also addressed this issue(Burrough,1998).The con-tribution of this paper is to describe how level set methods are:i)an appropriate scientific basis,and ii)able to perform stable time-space computations for modelling landscape pro-cesses.The level set method provides the following benefits:–it more directly models motion of spatial phenomena and may handle both expanding and contracting inter-faces,–is based upon differential equations related to the spatial dynamics of physical processes.Despite the potential for using level set methods in GIS and land-surface process modeling,there are no commercial or research systems that use this mercial sys-tems such as GRID(Gao et al.,1993),and research systems such as PCRaster(Wesseling et al.,1996)offerflexible andpowerful map algebra programming languages.But opera-tions that involve reaction-diffusive processing are specific to one context,such as groundwaterflow.We believe the level set method offers a more generic approach that allows a user to programflow and diffusive landscape processes for a variety of application contexts.We have shown that it pro-vides an appropriate theoretical underpinning and may be ef-ficiently implemented in a GIS.We have demonstrated its application for two landscape processes–albeit relatively simple examples–but these may be extended to deal with more complex and dynamic circumstances.The validation for improved environmental modeling tools ultimately rests in their uptake and usage by scientists and engineers.The tool may be accessed from the web site .au/projects/mapscript/(version with enhancements available April2005)for use with IDRSIS GIS(Eastman,1997)and in the future with ArcGIS. It is hoped that a larger community of users will make use of the methodology and implementation for a variety of environmental modeling applications.Edited by:P.Krause,S.Kralisch,and W.Fl¨u gelReviewed by:anonymous refereesReferencesArgent,R.:An Overview of Model Integration for Environmental Applications,Environmental Modelling and Software,19,219–234,2004.Bonham-Carter,G.F.:Geographic Information Systems for Geo-scientists,Elsevier Science Inc.,New York,1994. Burrough,P.A.:Dynamic Modelling and Geocomputation,in: Geocomputation:A Primer,edited by:Longley,P.A.,et al., Wiley,England,165-191,1998.Burrough,P.A.and McDonnell,R.:Principles of Geographic In-formation Systems,Oxford University Press,New York,1998. Gao,P.,Zhan,C.,and Menon,S.:An Overview of Cell-Based Mod-eling with GIS,in:Environmental Modeling with GIS,edited by: Goodchild,M.F.,et al.,Oxford University Press,325–331,1993.Goodchild,M.:A Geographer Looks at Spatial Information Theory, in:COSIT–Spatial Information Theory,edited by:Goos,G., Hertmanis,J.,and van Leeuwen,J.,LNCS2205,1–13,2001.Jørgensen,S.and Bendoricchio,G.:Fundamentals of Ecological Modelling,Elsevier,New York,2001.Julien,P.Y.,Saghafian,B.,and Ogden,F.:Raster-Based Hydro-logic Modelling of Spatially-Varied Surface Runoff,Water Re-sources Bulletin,31(3),523–536,1995.Moore,I.D.,Turner,A.,Wilson,J.,Jenson,S.,and Band,L.:GIS and Land-Surface-Subsurface Process Modeling,in:Environ-mental Modeling with GIS,edited by:Goodchild,M.F.,et al., Oxford University Press,New York,1993.Press,W.,Flannery,B.,Teukolsky,S.,and Vetterling,W.:Numeri-cal Recipes in C:The Art of Scientific Computing,2nd Ed.Cam-bridge University Press,Cambridge,1992.Pullar,D.:MapScript:A Map Algebra Programming Language Incorporating Neighborhood Analysis,GeoInformatica,5(2), 145–163,2001.Pullar,D.:Simulation Modelling Applied To Runoff Modelling Us-ing MapScript,Transactions in GIS,7(2),267–283,2003. Ritter,G.,Wilson,J.,and Davidson,J.:Image Algebra:An Overview,Computer Vision,Graphics,and Image Processing, 4,297–331,1990.Sethian,J.A.:Level Set Methods and Fast Marching Methods, Cambridge University Press,Cambridge,1999.Sklar,F.H.and Costanza,R.:The Development of Dynamic Spa-tial Models for Landscape Ecology:A Review and Progress,in: Quantitative Methods in Ecology,Springer-Verlag,New York, 239–288,1991.Sui,D.and R.Maggio:Integrating GIS with Hydrological Mod-eling:Practices,Problems,and Prospects,Computers,Environ-ment and Urban Systems,23(1),33–51,1999.Tilman,D.and P.Kareiva:Spatial Ecology:The Role of Space in Population Dynamics and Interspecific Interactions.Princeton University Press,Princeton,New Jersey,USA,1997. Wesseling C.G.,Karssenberg, D.,Burrough,P. A.,and van Deursen,W.P.:Integrating Dynamic Environmental Models in GIS:The Development of a Dynamic Modelling Language, Transactions in GIS,1(1),40–48,1996.。
韩拯:奋楫争先的“追光人”
韩拯:奋楫争先的“追光人”科学之友 402023-11一路高光是他“追光之旅”的名片韩拯是江苏扬州人,本科考入吉林大学物理学院,开始核物理专业学习。
之后考入中国科学院金属研究所材料学硕士专业。
2010年,他在法国攻读纳米电子学与纳米科技博士学位,导师对于他的评价是:“充满创新活力。
”此后,韩拯作为博士后在美国哥伦比亚大学物理系开展超高迁移率石墨烯-氮化硼异质结构的低温磁电性能方面的研究工作。
2015年10月博士后出站后,韩拯开展新型人工纳米器件的量子输运调控研究,这一研究领域涉及材料、电子、物理等基础研究学科,国际竞争激烈。
在韩拯回国研究期间,他多次与山西大学光电研究所的老师们进行密切的科研合作,在合作中了解到山西大学光电研究所扎实的学风和浓厚的科技底蕴。
2020年,他应邀出任山西大学光电研究所教目前世界上鳍片最薄的鳍式晶体管是中国制造,而它到底有多薄呢?鳍片宽度只有0.6纳米,相当于3个原子的厚度。
这项成果属于山西大学光电研究所教授、博士生导师韩拯,2020年底,他带领团队成员与湖南大学、中国科学院金属研究所等合作,最终成功制备出了世界上最薄的鳍式场效应晶体管。
关于“薄”的研究,就是韩拯的科研探索。
作为一名“85后”科研工作者,韩拯研究的,是尽可能找到二维世界里尺寸非常小的最薄功能材料,堆积成自然界中没有的新结构,探索其新奇有趣的物理性质,再利用这些物理现象来组装制造纳米尺度下的小型电子器件来服务于未来的应用。
多年来,每一次突破都是韩拯“追光之旅”的成果。
文|李炼授、博士生导师。
“科研是一个不断挑战未知领域、努力创新的过程。
科技教育则是在前人基础上使青年学子们了解已知领域,并热爱上挑战未知领域的过程。
”韩拯坦言,在山西省大力推进科技创新的大背景下,在彭堃墀院士创建的山西大学光电研究所这个平台上,青年科技工作者的发展空间十分广阔。
入职山西大学3年来,韩拯的科研成果多次入选科技领域重要进展,同时在山韩拯人物经纬创新 41西大学光电研究所聚集了一支优秀青年科学家队伍。
基于骨架树的荧光共焦图像神经树突棘自动分割与检测方法
Ap r a h fra tmae e me tt n a d d tcin o e d i sa d p o c o u o td s g nai n ee t fd n rt n o o e
s i e r m l oe c n e c n c li g a e n s eeo r e pn sfo fu r s e c o f a ma e b s d o k ltn te o
b t e n t e o gn p i t n et re on og t h nt l k lt n ew e h r i o n d t ag t i t e e i i a e eo .U e n mu s a nn e a e n w ih s h e i a h p t t i s s d mi i m p n i gt e b s d o eg ta e d — r t s r t n mo u e o e s eeo f e d t p n .I h s o n ai n h e e r h o ee t n o p n sfc s d o e c a g ci i d l ft k l tn o n r es i e n ti u d t ,t er s a c f tqi f i e u e n t h n e po h d i f o d o s o h o ev c o n l n s e on sa d t ev r c sd sa c fas ie ft e tra g ei e d p i t n h e t e i n eo p n .E p r n h wst a e p o o e p r a h c n e e — h i t x e me ts o t h r p s d a p o c a f c i h t t e ye ta tt e s ee o f h e d t n c iv etrr s l o p n sd t cin i l x r c h k lt n o e d n r e a d a h e e a b t e u t n s i e ee t . v t i e o
关于高三睡眠时间调查的英语作文
关于高三睡眠时间调查的英语作文全文共6篇示例,供读者参考篇1Sleep Time Survey of High School SeniorsHi everyone! My name is Timmy and I'm a 4th grader. My big sister Samantha is a senior in high school this year. She's been really stressed out lately with all her schoolwork, activities, and getting ready for college applications. My mom was worried that she hasn't been getting enough sleep, so she asked me to help out with a little project.For my school science fair last month, I did a survey about how much sleep high school seniors are getting each night. I asked Samantha and a bunch of her friends to record how many hours they slept every night for two weeks. Then I put all the data into charts and graphs to analyze it. It was a lot of work but I had fun learning about the scientific method!The ResultsHere are the key findings from my sleep survey of 25 high school seniors:On average, they only got 6 hours and 18 minutes of sleep per night during the school week. The recommended amount for teens is 8-10 hours!On weekends, they averaged 8 hours and 42 minutes of sleep, making up a bit for their Sleep deficit during the week.The least amount of sleep gotten on a school night was just 3 hours by Jessica R., who said she had 3 tests the next day.The most sleep anyone got was 11 hours on a Saturday by Kyle M., who told me "I was making up for a massive sleep debt."92% of the students said they feel occasionally or frequently sleep deprived due to homework, jobs, activities, etc.Causes of Sleep DeprivationWhen I interviewed the seniors, they gave me lots of reasons why they have trouble getting enough ZZZs:Homework overload was the #1 complaint. Samantha said, "I easily have 4-5 hours of homework every night, more if I have tests coming up. Sometimes I don't get to bed until 2 or 3 am."Extracurricular activities like sports, clubs, volunteering, and jobs also cut into sleep time. Tina told me, "Games and practicesgo until 7-8 pm some nights, then I have to rush home and start on my mountain of homework."Social media, TV, and video games are hard for some to resist at night instead of going to bed on time. As Max admitted, "I'll be working on homework then get distracted for hours on TikTok, YouTube, You name it."Stress, anxiety and caffeine consumption also make it tough to fall asleep at a reasonable hour. My friend Chloe is worried about her grades, test scores, getting into college, finances, you name it. She said, "My mind just races with everything I'm stressed about."ConsequencesNot surprisingly, all that missed sleep leads to some negative consequences for these busy high schoolers:Low energy, exhaustion and lack of focus were the top complaints. Isabella told me, "I'm a zombie by 3rd period. I can barely keep my eyes open in class."Poor eating habits and weight gain can also happen when you're too tired to motivate yourself to eat healthy. Pizza, energy drinks and candy were often mentioned.Moodiness, irritability and anxiety are common side effects of sleep deprivation. Samantha snapped at me the other day for no reason!Physical health issues like headaches, stomach aches, and getting sick more often can also result from lack of sleep.My RecommendationsAfter analyzing all my data and interviewing the seniors, I have some suggestions that could help high school students get more quality sleep:Have a strict lights-out time for phones, video games, etc. at night. The blue light really disrupts your sleep cycle.Avoid caffeine after 3 pm so it doesn't keep you wired at bedtime. Try a caffeinated drink like hot chocolate.Exercise regularly so you're physically tired out come bedtime. But avoid intense workouts too close to bedtime.Keep your bedroom cool, dark and quiet for optimal sleep conditions. Maybe use a fan, blackout curtains or white noise machine.Establish a calming pre-bed routine to let your body know it's sleep time soon. For example, take a warm bath/shower, read something boring, listen to calm music, etc.If you're staying up way too late doing homework, talk to your teachers. Maybe they can coordinate test dates better or reduce busywork assignments.Get outside for some natural sunlight every day, especially in the morning. Bright light helps reset your sleep-wake cycles.When you've just got too much on your plate, learn to prioritize and say no to some responsibilities for your ownwell-being.Well, that's my two cents! I hope Samantha and her friends start getting more shut-eye. Lack of sleep is no joke - it can really impact your health, grades and just overall functioning. A good night's sleep makes everything better. Thanks for reading my report, and don't forget to get your beauty sleep tonight!篇2A Big Look at How Much Sleep the Big Kids GetHi everyone! My name is Timmy and I'm a 3rd grader at Oakwood Elementary School. For my latest school project, I hadto do a big survey about sleep times for high school seniors. High schoolers are the really big kids who are almost adults. My teacher Mrs. Wilson helped me put together the questions and then I went around asking a bunch of seniors at the local high school.It was super interesting to learn about their sleep habits! I found out that most high school seniors don't get nearly enough sleep. The experts say teenagers need around 9 hours of sleep each night to be healthy and do well in school. But very few of the seniors I talked to were actually getting that much shuteye.Out of the 50 seniors I surveyed, only 6 of them said they regularly get 8 hours of sleep or more per night. That's just 12%! The vast majority were running on a lot less sleep than they're supposed to. 18 of the seniors (36%) told me they average 6-7 hours per night. But even more concerning, 24 of them (48%) said they only get 5 hours of sleep or less on most nights! Two seniors even admitted to running on a mere 3-4 hours quite often.Those last stats blew my mind. I can't even imagine trying to function on that little sleep as a little kid, let alone when you're a high school senior with hard classes, activities, sports, jobs, and getting ready for college. No wonder so many of them lookedtired with bags under their eyes! A few even dozed off while I was interviewing them.When I asked why they weren't getting enough sleep, I got a bunch of different reasons. Lots of them said they simply have too much homework and studying to do for all their tough classes. Others cited after-school activities like sports practices, clubs, jobs, or family responsibilities that keep them up late. Some said they do get their work done at a decent hour but then stay up way later than they should web browsing, watching shows and movies, gaming, or being on their phones. A few said they have troubles falling or staying asleep due to stress, anxiety, or overwhelming thoughts. And some admitted they just aren't disciplined enough with their sleep schedules and bedtimes.No matter the reasons, it's clearly a huge problem that so few high school seniors are getting the recommended amount of nightly sleep. I learned that not getting enough quality sleep can lead to difficulties concentrating, struggles staying awake in class, forgetfulness, moodiness, weight gain or loss, and even issues with mental health. Plus it makes kids way more likely to get sick or injured. Clearly this lack of sleep could really hurt a student's grades, test scores, extracurriculars, and overall quality of life. Yikes!Some of the seniors recognized how serious the issue is, but a bunch of them seemed to just accept their sleep deprivation as normal for high schoolers. They think running on fumes is just part of the deal when you're juggling so many responsibilities and preparing for college. While I suppose that might be true to some extent, I still think it's pretty unhealthy and concerning. Getting enough sleep should really be more of a priority.I asked the seniors if they had any tips for getting better sleep, and they shared some good advice. Many said it's crucial to find a consistent sleep schedule that allows for 8-9 hours per night and stick to it, even on weekends. No more all-nighters cramming for tests or procrastinating on big projects until the last minute! They recommend doing homework right after school, taking breaks to recharge, and then having leisure time in the evenings once all work is complete. Avoid too much screen time before bed and develop a calming pre-bed routine like reading, meditating, or light stretching.Some seniors also suggested blocking out set times for sports, activities, and a job each week rather than letting those bleed into homework and sleep time. If possible, cutting back on extracurriculars can provide more sleep opportunities too. Getting regular exercise and making time to hang out withfriends were also cited as sleep boosters. And eating a healthy diet full of fruits, veggies, and lean protein can enhance sleep quality as well.Overall, I learned that the sleep deprivation situation for high school seniors is pretty troubling. Really hardly any of them are getting the recommended 9 hours on a regular basis. I'm glad I got to dig deeper into understanding the issue and hearing directly from seniors about the impacts, causes, and advice for getting better sleep. Hopefully when I'm their age, I'll be smarter about making sleep more of a priority! Thanks for reading my report, everybody. Time for me to go take a nap!篇3My Big Sleep SurveyHi everyone! My name is Timmy and I'm in 3rd grade. I love school, playing outside, and learning about science and nature. This year for my science fair project, I decided to do a really big survey all about how much sleep high school students get every night. Why did I pick that topic? Well, my teenage brother Jack is always complaining about being tired, and I wanted to find out if he's the only one who doesn't get enough sleep or if it's a bigger problem.To do my survey, I made a questionnaire with questions like "What time do you go to bed on school nights?" and "What time do you wake up in the morning?" and "Do you take naps after school?". I also asked them to tell me how many hours of sleep they get each night. I brought the surveys to the high school near my house and had the students fill them out during their lunch periods over a few days.I was really surprised by the results! Out of the 237 students who took my survey, a huge number of them - 183 to be exact - said they get less than the recommended 8-10 hours of sleep per night for teens. The average amount of sleep they reported was only 6 hours and 43 minutes! No wonder Jack is always so cranky.The reasons they gave for not getting enough sleep were all pretty similar. Lots of them said they have too much homework and don't have time to get to bed until really late after finishing all their assignments. Others said they have after-school activities like sports practices or clubs that keep them busy until night time. And a bunch of them admitted they just tend to stay up way too late watching TV, playing video games, or being on their phones in bed. Teenagers sure do love their screens!I was really interested to learn more about why sleep is so important, so I did a bunch of extra research on it. Sleeping is how our bodies and brains recharge from all the hard work they do during the day. Not getting enough quality sleep can make it harder to concentrate, remember things we learned, and perform our best. It can also put teenagers at greater risk for weight gain, injuries, illness, and feeling depressed or anxious. Yikes!Based on my findings, I have a few suggestions forteenagers on how to get more sleep:Develop a consistent bedtime routine and stick to it every night, even on weekends. Things like taking a warm shower/bath, reading for pleasure for 20-30 mins, and winding down with some light stretching before bed can all help signal to your body that it's time to sleep.Stop using electronics at least an hour before trying to fall asleep. The bright screens from TVs, phones, tablets, video games, etc can trick your brain into thinking it's still daytime which makes it harder to sleep.Find ways to manage stress from schoolwork, activities, social life, etc. Don't leave everything until the last minute!Things like daily exercise, talking to friends/family, journaling, meditating, and taking breaks can all help you destress.If possible, ask about changing your school's start time to a bit later in the morning. Teenagers' circadian rhythms (aka their natural sleep/wake cycles) are different and they tend to get sleepy later at night. An 8:30am start may be better than a7:30am one.If nothing else works, take a 20-30 minute nap after school to reenergize but not so long that it disrupts your night sleep. Just be sure to nap somewhere quiet and peaceful.I really hope high schoolers like my brother Jack will follow some of these tips so they aren't so tired and grumpy anymore! Getting enough sleep is one of the most important things we can do for our health and happiness. And if teenagers still have trouble, they should talk to their parents or doctor to make sure nothing else is keeping them from getting quality zzzs.Well, thanks for reading about my big sleep survey results! Even though I'm just a kid, I think projects like this show that young students can do some pretty cool and serious research when we put our minds to it. I'm already looking forward to picking my next science fair topic. Maybe I'll survey college kidsnext to see if they've learned to make sleep more of a priority or not. Until then, SLEEP TIGHT EVERYONE!篇4My Big Sister's Sleep SurveyMy big sister Jessica is in her last year of high school, which is called senior year. She has to do a big research project to graduate. For her project, she decided to study how much sleep her classmates get each night. I think that's a pretty boring topic, but she says it's really important for health and school performance.Jessica explained to me that high school seniors have a lot of homework, activities like sports or clubs, jobs, and also have to study for big tests to get into college. Because of all that, many of them don't get enough sleep. Not sleeping well can make you feel grumpy, have trouble concentrating, and even get sick more often.To start her research, Jessica made a survey with questions about what time kids go to bed, what time they wake up, and other stuff related to sleep habits. She passed it out to all the seniors at her school during their homeroom period. Over 200 students filled it out, which she said gives her good data.After collecting all the surveys back, Jessica entered all the answers into a spreadsheet program on her computer. She used formulas to calculate things like the average number of hours each person slept per night. Jessica worked really hard organizing and analyzing all that data!Here are some of the interesting findings from Jessica's sleep survey:The average bedtime for seniors was 12:32am on school nights. On weekends, the average bedtime was 1:45am.Wake-up times were much earlier on school days, with the average being 6:12am versus 9:37am on weekends.Based on those sleep and wake times, the average senior only got 5 hours and 40 minutes of sleep on school nights! On weekend nights it was more like 8 hours.Over 75% of students reported feelings of sleepiness or struggling to stay awake during the school day.Students who worked at jobs after school tended to get even less sleep, often only 5 hours or less.Only 8% of seniors said they slept 8 hours or more per school night, which is the recommended amount for teens.Jessica showed me all these numbers and percentages in charts and graphs, which was kind of confusing to me. But she explained what the pictures and data meant in a way I could understand.Jessica's conclusions from the survey were that most high school seniors at her school are seriously sleep deprived, especially on school nights. She thinks schools should start later in the morning because teenagers naturally prefer going to bed late and sleeping late. Jessica is going to propose some solutions in her final project, like limiting homework, no weekend jobs, and education for students on good sleep habits.I'm glad Jessica chose a topic that is important for student health and success, even if it seems a little boring at first. Her survey showed that lack of sleep is a real problem many high schoolers face. Maybe if schools make some changes based on Jessica's findings, students can get more sleep and have more energy to learn.I just hope Jessica is getting enough sleep herself while working so hard on this big project! I'm proud of my big sister for doing such a good job. Once she finishes and gets her diploma, she's going to teach me all about how to get a goodnight's sleep when I'm a busy high school student myself someday.篇5My Big Survey on Sleep for High School SeniorsHi everyone! My name is Tommy and I'm in 5th grade. For my school science project this year, I decided to do a big survey about how much sleep high school seniors get each night. Seniors are kids who are almost done with high school and will be going to college or getting jobs soon. I picked this topic because I've heard that a lot of seniors don't get enough sleep, and not getting enough sleep can make you tired, grumpy, and have trouble paying attention in class.My survey had questions like:What time do you usually go to bed on school nights?What time do you usually wake up on school days?How many hours of sleep do you get per night?Do you feel well-rested when you wake up in the morning?Do you ever take naps after school?I made the survey online using a website called SurveyMonkey. It was pretty easy and fun to make! Then I shared the link to my survey with the seniors at my local high school. I also printed out some paper copies and went around asking seniors to fill it out during their lunch periods.In total, I got 187 seniors to complete my survey! I was really happy with that number. My science teacher Mr. Kozlowski showed me how to analyze all the data using spreadsheets. Here's what I found out:Sleep DurationThe average senior reported getting only 6 hours and 18 minutes of sleep per night! That's way less than the 8-10 hours that experts recommend for teenagers. Only 12% of the seniors said they get 8 hours of sleep per night. A whopping 71% get less than 7 hours! The lowest amount reported was just 4 hours per night from one very sleep-deprived senior.BedtimesMost seniors go to bed pretty late. The average bedtime reported was 12:24am! Just a small number (14%) said they go to bed before 11pm on school nights. The latest bedtime was 3am from one student who must be a really big night owl.Wake Up TimesOn the flip side, seniors have to get up pretty early for school. The average wake up time was 6:32am, which doesn't seem too crazy. But some students reported waking up as early as 5am! No thanks! A small group of lucky seniors (8%) don't have to wake up until after 8am.Feeling RestedWhen I asked "Do you feel well-rested when you wake up in the morning?", a huge 82% of seniors said NO! They drag themselves out of bed feeling exhausted almost every day. Only 18% said they usually feel rested. That's really bad!NapsTo try and make up for their sleep debt, a lot of seniors told me they take naps after school. Around 44% said they nap for1-2 hours pretty often, and another 21% said they'll nap if they have time or feel really tired that day. The other 35% said they never or rarely nap.ConclusionsAfter looking at all the data, I can say that seniors definitely don't get enough sleep! They go to bed way too late at night, don't get the recommended 8-10 hours, and feel tired almostevery morning as a result. A lot of them try to catch up by napping, but naps aren't as good as just getting great sleep at night.There could be lots of reasons why seniors sleep so little. Maybe they have piles of homework, participate in sports or clubs after school, have jobs, or just prefer to stay up late browsing their phones and social media. Whatever the reasons, not getting enough zzz's is bad for their health and could make it hard to do well in school.If I were a high school senior, here's what I would do to get more sleep:Go to bed earlier, like 10 or 11pm every nightStop looking at screens at least 1 hour before bedtimeSleep in a cool, dark, quiet bedroomDon't consume caffeine or sugar late in the dayTake short naps if I feel tired, but not too late in the eveningStick to a consistent sleep schedule, even on weekendsSeniors, please try to get more sleep! Your body and brain need it to grow and recharge. You'll feel so much better andhave more energy to slay your final year of high school. Let me know if you have any other questions!That's all for my big sleep survey report. I had a lot of fun doing this project and learned that even older kids struggle to get enough sleep. Thanks for reading, sleep tight, and sweet dreams!篇6Sleep Survey for High School SeniorsHi there! My name is Timmy and I'm in 4th grade. I love learning about science and doing experiments. Recently, my big sister Anna who is a senior in high school had to do a science project for her class. She decided to do a survey about how much sleep high school seniors get each night. I thought it was a really cool project, so I helped her out!Anna made a survey with questions like "What time do you typically go to bed on a school night?" and "What time do you wake up in the morning?" She also asked about napping, weekends, and if people felt well-rested. Then she shared the survey link with all the seniors at our high school.We got over 200 responses which was awesome! Anna and I looked through all the data and made some interesting discoveries. Most seniors said they go to bed around 11pm or midnight on school nights. That seems so late to me - I'm usually in bed by 8:30pm! A lot of them said they struggle to fall asleep before midnight too because they have so much homework.The wake-up times for seniors varied a lot more. Some said they wake up as early as 5:30am, while others could sleep until 7:30 or even 8am. I wake up at 6:45am, so I'm somewhere in the middle for high schoolers. We calculated that on average, the seniors get around 6-7 hours of sleep per night.When we asked if they feel well-rested, a huge number - almost 75% - said no! Many wrote comments saying they are constantly tired and have trouble staying awake in class. Some even said they take naps after school. Just imagining being that tired all the time makes me feel exhausted!On the weekends, it was a totally different story. Most seniors reported sleeping 9-11 hours per night when they don't have school the next day. I was surprised it was that much! Anna said they are probably "catching up" on all the missed sleep from the school week.After looking at all the data, Anna and I had a lively discussion about why high school seniors seem to be so sleep deprived. We think there are a few main reasons:Homework load - Seniors take a ton of tough classes with lots of homework, projects, and studying. They have to stay up late to get it all done.Extracurriculars - Many seniors are involved in sports, clubs, jobs, etc. after school. Their days are just packed.Social life - Hanging out with friends, using phones/screens late at night, and just being a typical teenager makes it hard to get enough sleep.Bad habits - Some seniors admit to binge-watching shows, playing video games, or scrolling social media instead of going to bed on time.Stress - The pressure of grades, college applications, and general life stress makes it difficult for seniors to fall asleep at a reasonable hour.Anna's conclusion is that based on all the data, high school seniors are probably not getting the recommended 8-10 hours of sleep per night that teenagers need. She thinks schools shouldstart later so seniors don't have to wake up as early. I agree - who wants to be sleepy all day at school?!Personally, I'm glad I'm still in elementary school and don't have to worry about all that yet. I get my beauty sleep every night! Although I have to admit, staying up late to watch movies or play video games does sound kind of fun sometimes. But I'll stick to getting my zzz's for now while I can. Being a kid rules!Anna's project taught me a lot about the importance of sleep, especially for hard-working high school students. I just hope by the time I'm a senior, things are different and we can all get the rest we need. School is hard enough without feeling like a zombie!That's my take on Anna's sleep survey. Let me know if you have any other questions! I'll be sure to get a full 8 hours of sleep tonight. Nighty night!。
2022年02月上海交通大学医学院精准医学研究院曾汉林课题组诚聘肿瘤研究助理研究员及博士后笔试历年高
2022年02月上海交通大学医学院精准医学研究院曾汉林课题组诚聘肿瘤研究助理研究员及博士后笔试历年高频考点试题库集锦答案解析全文为Word可编辑,若为PDF皆为盗版,请谨慎购买!卷I一.单选题(共25题)1.十八大以来的五年,是党和国家发展进程中极不平凡的五年。
五年来的成就是()。
A.深层次的、根本性的B.全覆盖的、压倒性的C.多层次的、立体性的D.全方位的、开创性的答案:D本题解析:暂无解析2.下列()不是我国教育权法律救济体制中的制度的一种。
A.学生申诉制度B.行政复议制度C.行政诉讼制度D.司法调解制度答案:D本题解析:我国教育权法律救济体制主要由以下各项制度构成:教师申诉制度、学生申诉制度、行政复议制度、行政诉讼制度、行政赔偿制度和民事诉讼制度。
3.新时代党的教育事业发展的根本任务是()。
A.立德树人B.素质教育C.教育公平D.教育改革答案:A本题解析:暂无解析4.根据我国民法通则的规定,学校及其他教育机构应属于()。
A.企业法人全文为Word可编辑,若为PDF皆为盗版,请谨慎购买!B.社会团体法人C.机关法人D.事业单位法人答案:D本题解析:暂无解析5.山东真题:根据我国教育法规体系的横向结构,下列选项中属于基础教育法的是()。
A.《中华人民共和国义务教育法》B.《中华人民共和国高等教育法》C.《中华人民共和国教育法》D.《中华人民共和固职业教育法》答案:A本题解析:已颁布施行的《中华人民共和国义务教育法》为基础教育法的一个组成部分。
6.高等教育行政行为一经作出,即被推定为合法有效的,其约束力随之产生。
这反映高等教育行政行为的特征是()。
A.从属法律性B.效力先定性C.强制性D.单方意志性答案:B本题解析:暂无解析7.取消考试资格是教育行政处罚中的()。
A.申诫罚B.能力罚C.财产罚D.人身罚答案:B本题解析:暂无解析8.根据《高等教育法》,高等学校的具体标准由谁制定()A.人大常委会B.国务院C.教育部D.各省、自治区、直辖市人民政府答案:B本题解析:暂无解析全文为Word可编辑,若为PDF皆为盗版,请谨慎购买!9.法学理论上讲的教育法律事实是指能够引起教育法律关系产生、变更或消失的客观情况,它可分为两类:一是教育法律事件,二是()。
群体智能优化算法-萤火虫算法
第八章萤火虫算法8.1介绍萤火虫(firefly)种类繁多,主要分布在热带地区。
大多数萤火虫在短时间内产生有节奏的闪光。
这种闪光是由于生物发光的一种化学反应,萤火虫的闪光模式因种类而异。
萤火虫算法(FA)是基于萤火虫的闪光行为,它是一种用于全局优化问题的智能随机算法,由Yang Xin-She(2009)[1]提出。
萤火虫通过下腹的一种化学反应-生物发光(bioluminescence)发光。
这种生物发光是萤火虫求偶仪式的重要组成部分,也是雄性萤火虫和雌性萤火虫交流的主要媒介,发出光也可用来引诱配偶或猎物,同时这种闪光也有助于保护萤火虫的领地,并警告捕食者远离栖息地。
在FA中,认为所有的萤火虫都是雌雄同体的,无论性别如何,它们都互相吸引。
该算法的建立基于两个关键的概念:发出的光的强度和两个萤火虫之间产生的吸引力的程度。
8.2天然萤火虫的行为天然萤火虫在寻找猎物、吸引配偶和保护领地时表现出惊人的闪光行为,萤火虫大多生活在热带环境中。
一般来说,它们产生冷光,如绿色、黄色或淡红色。
萤火虫的吸引力取决于它的光照强度,对于任何一对萤火虫来说,较亮的萤火虫会吸引另一只萤火虫。
所以,亮度较低的个体移向较亮的个体,同时光的亮度随着距离的增加而降低。
萤火虫的闪光模式可能因物种而异,在一些萤火虫物种中,雌性会利用这种现象猎食其他物种;有些萤火虫在一大群萤火虫中表现出同步闪光的行为来吸引猎物,雌萤火虫从静止的位置观察雄萤火虫发出的闪光,在发现一个感兴趣趣的闪光后,雌性萤火虫会做出反应,发出闪光,求偶仪式就这样开始了。
一些雌性萤火虫会产生其他种类萤火虫的闪光模式,来诱捕雄性萤火虫并吃掉它们。
8.3萤火虫算法萤火虫算法模拟了萤火虫的自然现象。
真实的萤火虫自然地呈现出一种离散的闪烁模式,而萤火虫算法假设它们总是在发光。
为了模拟萤火虫的这种闪烁行为,Yang Xin-She提出了了三条规则(Yang,2009)[1]:1. 假设所有萤火虫都是雌雄同体的,因此一只萤火虫可能会被其他任何萤火虫吸引。
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
NEARLY–LIGHT CYCLES IN EMBEDDED GRAPHS ANDCROSSING–CRITICAL GRAPHSMARIO LOMEL´I AND GELASIO SALAZARAbstract.Wefind a lower bound for the proportion of face boundaries of anembedded graph that are nearly–light(that is,they have bounded length and atmost one vertex of large degree).As an application,we show that every sufficientlylarge k–crossing–critical graph has crossing number at most2k+23.1.IntroductionIt is quite natural to inquire about the existence of“light”subgraphs in a given family G of graphs.Recall that if H is a subgraph of G,then the weight w(H)of H in G is the sum of the valences in G of the vertices in H.If there is an integer w such that every graph G in G that contains a subgraph isomorphic to H contains one such subgraph with weight at most w in G,then H is light in G.Most research on light subgraphs has focused on the case in which G is a family of graphs embedded in some compact surface(see for instance[1,2,6,7,8,9,12,13].Although under certain conditions one can guarantee the existence of light cycles in embedded graphs(see[5]),this is not always the case:every cycle in a wheel either contains a hub vertex(which can have arbitrarily high degree),or is arbitrarily long (as long as the degree of the hub).In view of this,a natural way to proceed in this context is to inquire about the existence of“nearly–light”cycles.Let ,∆be positive numbers.A cycle C in a graph G is( ,∆)–nearly–light if the length of C is at most ,and at most one vertex of C has degree greater than∆.If G is embedded,we define an( ,∆)–nearly–light face boundary similarly,with the convention that an edge that is traversed twice in the boundary walk of a face contributes in two to the length of that face boundary.In[14],Richter and Thomassen investigated the existence of nearly–light cycles, and proved that every planar graph has at least one(5,11)–nearly–light face bound-ary.One of the aims in this work is to refine this statement,and show that plane (moreover,embedded)graphs have not one but many nearly–light face boundaries.Date:February10,2006.1991Mathematics Subject Classification.05C10,05C62.Key words and phrases.Light,nearly–light,graph embedding,crossing number,crossing–critical.Supported by FAI–UASLP and by CONACYT through Grant J32168E.12MARIO LOMEL ´I AND GELASIO SALAZARTheorem 1.Let 0<ε<1/6,and let G be a simple connected graph with mini-mum degree at least 3,embedded in a (orientable or nonorientable)surface of Euler characteristic χ.Let F (G )denote the set of faces of G .Then G contains at least (2χ−1)+ 14−3ε2 |F (G )|face boundaries that are (6,2/ε)–nearly–light.The problem of the existence of nearly–light cycles is raised and attacked in [14]in the context of crossing–critical graphs.We recall that the crossing number cr(G )of a graph G is the minimum number of pairwise crossings of edges in a drawing of G in the plane.A graph G is k –crossing–critical if its crossing number is at least k ,but cr(G −e )<k for every edge e of G .In [14],the existence of a nearly–light cycle is used to prove that every k –crossing–critical graph has crossing number at most 2.5k +16.As we show below,Theorem 1implies the following statement on the crossing numbers of sufficiently large crossing–critical graphs.Theorem 2.For each k >0there is an n (k )with the following property.If G is a k –crossing–critical graph with at least n (k )vertices of degree greater than two,then cr(G )≤2k +23.We note that the condition in this statement on the degrees of the vertices (greater than two)is unavoidable,since subdivisions of edges change neither the crossing number of a graph nor its criticality.Crossing–critical graphs are objects of natural interest whose structure has been the object of recent investigations (see for instance [4]).Moreover,upper bounds for the crossing number of crossing–critical graphs also have an important application.Indeed,as Richter and Thomassen observed,their bound cr(G )≤2.5k +16for k –crossing–critical graphs implies that if H is an arbitrary graph with cr(H )=k ,then there is an edge e in H such that cr(H −e )≥(2k −37)/5.Along the same lines,it is readily checked that our Theorem 2implies the following.Corollary 3.For each k >0there is an n (k )with the following property.If H has at least n (k )vertices of degree greater than two,and cr(H )=k ,then H has an edge e such that cr(H −e )≥(k −26)/2. Very recently,Fox and Toth improved this result for sufficiently dense graphs [3].We prove Theorems 1and 2in Sections 2and 3,respectively.2.Nearly–light face boundaries in embedded graphsIn this section we show that the technique used in the proof of Theorem 1in [14]can be refined to give a proof of Theorem 1.For an embedded graph G ,we let V (G ),E (G ),and F (G )denote the sets of vertices,edges,and faces of G ,respectively.Proof of Theorem 1.As in [14],for each face f of G let the weight w (f )be the sum v ∼f (1/d (v )),where d (v )denotes the degree of vertex v and v ∼f means that v isNEARLY–LIGHT CYCLES AND CROSSING–CRITICAL GRAPHS 3incident with f .Thus,for each face f ,w (f )≤l (f )/3,where l (f )denotes the length of the boundary of f .As in the proof of Theorem 1in [14],we note that f ∈F (G )w (f )=|V (G )|,and f ∈F (G )l (f )=2|E (G )|.Thus,Euler’s formula implies that f {w (f )−l (f )/2+1}≥χ(note that strict inequality may occur,as there may be a noncontractible curve in the surface that does not intersect the embedded graph G ).Let us say that a face f is good if w (f )−l (f )/2+1>−1/6+ε.We complete the proof by showing that the following statements hold.(A)For each good face f ,the face boundary of f is (6,2/ε)–nearly–light.(B)There are at least (2χ−1)+ 1/4−3ε/2 |F (G )|good faces.Let f be a good face,and suppose that l (f )>6.Since −1/6+ε<w (f )−l (f )/2+1,and w (f )≤l (f )/3,then −1/6+ε<−l (f )/6+1≤−7/6+1=−1/6,contradicting the assumption ε>0.Thus l (f )≤6.Now suppose that at least two vertices v incident with f have d (v )>2/ε.Therefore w (f )<(l (f )−2)/3+2(ε/2)=(l (f )−2)/3+ε.Since −1/6+ε<w (f )−l (f )/2+1,it follows that −1/6+ε<l (f )/3−2/3+ε−l (f )/2+1=−l (f )/6+1/3+ε.Hence l (f )<3,contradicting the assumption that G is simple.Hence at most one vertex incident with f has degree greater than 2/ε.This proves (A).Let D (G )denote the set of good faces.Now f ∈D (G ){w (f )−l (f )/2+1}+ f ∈(F (G )\D (G )){w (f )−l (f )/2+1}≥χ.By definition,each f ∈(F (G )\D (G ))satisfies w (f )−l (f )/2+1≤−1/6+ε.On the other hand,every face f has w (f )−l (f )/2+1≤1/2.Thus |D (G )|/2+(|F (G )|−|D (G )|)(−1/6+ε)≥χ.An easy manipulation then yields that |D (G )|> (1/6)−ε(2/3)−ε |F (G )|+χ/(2/3−ε).Hence |D (G )|> 1/4−3ε/2 |F (G )|+χ/(2/3−ε).We finally note that 0<ε<1/6implies that,if χ≤0,then χ/(2/3−ε)≥2χ>2χ−1.On the other hand,if χ>0then χ=1or 2,and so χ>0implies χ/(2/3−ε)>2χ−1.It follows that regardless of the sign of χ,χ/(2/3−ε)>2χ−1.Therefore |D (G )|> 1/4−3ε/2 |F (G )|+(2χ−1).This proves (B).3.Crossing–critical graphsIn this section we prove Theorem 2.The proof has two main ingredients.First we show (Lemma 4)that large crossing–critical graphs have (6,12)–nearly–light cycles.Then we invoke a result (Lemma 5)whose proof is implicit in the proof of Theorem 3in [14],namely that the existence of a nearly–light cycle in a crossing–critical graph yields an upper bound for the crossing number of the graph.Lemma 4.For each integer k >0,there is an n (k )with the following property.Let G be a simple k –crossing–critical graph with minimum degree at least 3.Suppose that |V (G )|≥n (k ).Then G contains a (6,12)–nearly–light cycle.4MARIO LOMEL ´I AND GELASIO SALAZARProof.First we observe that if G is k –crossing–critical,then G can be embedded in the orientable surface Σk of genus k (that is,Euler characteristic χ=2−2k ).This follows since G contains a set of at most k edges whose deletion leaves G planar.We show that this embedding has a (6,12)–nearly–light face boundary.This com-pletes the proof,as this face boundary contains the required (6,12)–nearly–light cycle.Apply Theorem 1to G embedded in Σk ,with ε=4/25.This yields the existence of at least (2χ−1)+(1/4−6/25)|F (G )|=(3−4k )+(1/4−6/25)|F (G )|face bound-aries that are (6,12)–nearly–light (note that a (6,12.5)–nearly–light face boundary is (6,12)–nearly–light).We finally note that if |V (G )|is sufficiently large (compared to k ),then (by Euler’s formula)so is |F (G )|,and this in turn guarantees that (3−4k )+(1/4−6/25)|F (G )|≥1.Therefore,if |V (G )|is sufficiently large,then there is a (6,12)–nearly–light face boundary. The proof of the first inequality in the following lemma is implicit in the proof of Theorem 3in [14].The second inequality follows from the first inequality and the definition of an ( ,∆)–nearly–light cycle.Lemma 5.Let G be a k –crossing–critical graph,and let s >0.Suppose that G has a cycle C with a vertex v such that u ∈C \{v }(d (u )−2)≤s .Thencr(G )≤2(k −1)+s/2.Thus,if G has an ( ,∆)–nearly–light cycle,thencr(G )≤2(k −1)+(∆−2)( −1)2.Proof of Theorem 2.Let G be a k –crossing–critical graph.By supressing vertices of degree two if necessary (this affects neither the crossing number nor the criticality)we may assume that G has no vertices of degree two or less.Now suppose that |V (G )|≥n (k ),where n (k )is as in Lemma 4.As in the proof of Theorem 3in [14],we can assume that G is simple,as otherwise cr(G )≤2(k −1),in which case we are done.Lemma 4then applies,and yields the existence of a (6,12)–nearly–light cycle in G .By applying Lemma 5we obtain cr(G )≤2(k −1)+(10)(5)/2=2k +23.4.Concluding RemarksIt is natural to inquire about the tightness of the bound in Theorem 1.How much can the coefficient of |F (G )|be improved by allowing larger values of and ∆?Consider the following construction.Let H 0be a graph isomorphic to K 4,4−e ,and let u,v denote the degree 2vertices of H 0.Now let G n be obtained by taking n copies of H 0,and identifying them along u and v .Thus G n has two vertices of degree 2n ,and 2n vertices of degree 3.Moreover,every planar embedding of G n hasNEARLY–LIGHT CYCLES AND CROSSING–CRITICAL GRAPHS5 n faces(of size four)incident with both u and v,and2n faces(of size three)incident with two degree3vertices and exactly one copy of H0.Thus,for everyfixed∆,if n is sufficiently large then exactly two thirds of the faces of any embedding of G n are ( ,∆)–nearly–light.This shows that the coefficient of|F(G)|in Theorem1cannot be improved to a value greater than2/3,regardless of the size of∆.On the other hand,the upper bound2/3on the coefficient of|F(G)|can be almost attained as a lower bound,as the following statement claims.Theorem6.For eachε>0and integerχ≤2there exist 0:= 0(ε,χ),∆0:=∆0(ε,χ),c:=c(ε,χ)with the following property.Let G=(V,E)be a simple con-nected graph with minimum degree at least3,embedded in a surface with Euler characteristicχ.Let F denote the set of faces of G.Then G contains at least2 3−ε|F|+c face boundaries that are( 0,∆0)–nearly–light.This result can be proved by direct geometrical methods(see[11]).Unfortunately,these arguments are not nearly as neat and elegant as the powerful technique,intro-duced by Lebesgue in[10],that we used in the proof of Theorem1.References[1]I.Fabrici,On vertex-degree restricted subgraphs in polyhedral groups,Discrete Math.256(2002),no.1-2,105–114.[2]I.Fabrici,E.Hexel,S.Jendrol’,and H.Walther,On vertex–degree restricted paths in polyhedralgraphs,Discrete Math.212(2000),no.1-2,61–73.[3]J.Fox and C.Toth,The Decay of Crossing Numbers.Manuscript(2005).[4]P.Hlinˇe n´y,Crossing-Number Critical Graphs have Bounded Pathwidth,bin.Theory Ser.B88(2003),347–367.[5]S.Jendrol’,T.Madaras,R.Sotk,and Z.Tuza,On light cycles in plane triangulations,DiscreteMath.197/198(1999),453–467.[6]S.Jendrol’and P.J.Owens,On light graphs in3–connected plane graphs without triangular orquadrangular bin.17(2001),no.4,659–680.[7]S.Jendrol’and H.–J.Voss,Light paths with an odd number of vertices in large polyhedral maps,b.2(1998),No.4,313–324.[8]S.Jendrol’and H.–J.Voss,Subgraphs with restricted degrees of their vertices in polyhedral mapson compact2–b.5(2001),no.2.,211–226.[9]S.Jendrol’and H.–J.Voss,Light paths in large polyhedral maps with prescribed minimum degree,bin.25(2002),79–102.[10]H.Lebesgue,Quelques cons´e quences simples de la formule d’Euler,J.Math.19(1940),27–43.[11]M.Lomel´ıand G.Salazar,Addendum to“Nearly–light cycles and crossing–critical graphs”.Manuscript available at http://www.ifisica.uaslp.mx/∼gsalazar.[12]B.Mohar,Light paths in4-connected graphs in the plane and other surfaces,J.Graph Theory34(2000),No.2,170–179.[13]B.Mohar,R.ˇSkrekovski,H.–J.Voss,Light subgraphs in planar graphs of minimum degree4and edge–degree9,J.Graph Theory44(2003),No.4,261–295.[14]R.B.Richter and C.Thomassen,Minimal graphs with crossing number at least k,-bin.Theory Ser B58,No.2(1993),217–224.6MARIO LOMEL´I AND GELASIO SALAZARFacultad de Ciencias,UASLP,San Luis Potosi,SLP.Mexico78000 Instituto de F´ısica,UASLP,San Luis Potosi,SLP.Mexico78000 E-mail address:gsalazar@ifisica.uaslp.mxURL:http://www.ifisica.uaslp.mx/∼gsalazar。