A generalized FFT algorithm on transputers

合集下载

基于NT降阶算法的区间二型模糊系统辨识

基于NT降阶算法的区间二型模糊系统辨识王哲【摘要】Due to the defects in description system uncertainty of the traditional T-S fuzzy description system, type-2 T-S fuzzy system has received extensive attention. In according with the low efficiency of common type reduction algorithm for interval two type fuzzy set, the NT type reduction algorithm was used for interval type-2 fuzzy system identification. The NT type reduction algorithm avoid the complexity iterative search operation, directly using the upper and lower bounds of first membership function, and then directly get the results of the fuzzy system. The simulations result shows that NT type reduction algorithm can improve identification efficiently without reduce identification accuracy.%由于传统T-S模糊描述系统不确定性方面的缺陷,二型T-S模糊系统得到了广泛关注.针对常见区间二型模糊集合的降阶算法存在的效率低下的问题,本文利用NT降阶算法进行区间二型模糊系统的辨识.NT降阶算法避免了复杂的迭代搜索操作,直接利用首隶属度函数的上、下限进行降阶运算,然后直接得到解模糊化结果.仿真实例表明,利用NT降阶算法能够在不降低辨识精度的情况下,提高辨识效率.【期刊名称】《仪器仪表用户》【年(卷),期】2018(025)006【总页数】4页(P17-20)【关键词】区间二型模糊集合;降阶;T-S模糊系统;模糊辨识;NT降阶算法【作者】王哲【作者单位】天津现代职业技术学院,天津 300350【正文语种】中文【中图分类】TP273+.40 引言近些年，T-S模糊模型在非线性系统辨识方面取得了很好的效果。

基于鸽群算法的Fuch混沌蝗虫算法

和后期都是以相同的方式进行更新，不能满足搜索过程中的实
际情况，因为在迭代前期权重ｃ需要增大到
一个较小权值，这样易于局部开发，避免了算法陷入局部最优，
提高了收敛速度，平衡了全局与局部搜索性能。所以针对ｃ的
缺点，通过改进文献［１５］的递变函数来对权重因子ｃ进行优化
速朝着最优目标靠近，以提高收敛速度，且较大的ｃ值能够对
全局进行充分搜索；在迭代后期，权重ｃ缓慢下降，直到迭代次
数达到最大时ｃ值降到ｃｍｉｎ，可以使蝗虫对周围空间进行充分探索，并且较小的ｃ值使当前的蝗虫最优个体离目标较近易于
局部开发，不仅提高了全局探索能力而且避免了前期过早陷入
局部最优。通过非线性递减权重来动态调整ＰＦＧＯＡ全局与局
２４正余弦鸽群混合策略
传统蝗虫算法使用式（４）来进行蝗虫最优位置更新，蝗虫
种群中的个体都直接相互作用，并且个体可以通过其他蝗虫的
位置来确定移动的方向。蝗虫之间存在排斥力和吸引力，使蝗
虫在搜索空间中进行探索，蝗虫可以通过社交作用来进行信息
交流，加快收敛速度。但是在迭代后期蝗虫个体之间的信息交
流会减弱，很容易陷入局部最优，而无法获得全局最优解。于是采用正余弦算法的［１６］分段思想来对蝗虫位置进行更新，使
迭代次数。
２２非线性惯性权重递减策略
在传统蝗虫算法中，式（４）中的外部ｃ类似于粒子群算法
的惯性权重，用于调整目标周围的搜索性能，内部ｃ可以调节吸引区、舒适区和排斥区之间的相互作用［１４］。内部ｃ和外部ｃ
的公式相同，是平衡探索与开发能力的关键，它以线性方式随
着迭代次数的增加而减少。但是，所有蝗虫的位置在迭代前期
２ＰＦＧＯＡ
２１Ｆｕｃｈ混沌初始化

机动目标距离徙动校正与检测算法

机动目标距离徙动校正与检测算法贺雄鹏;廖桂生;许京伟;黄鹏辉;魏嘉琪【摘要】在机动目标的检测中,目标的速度、加速度会产生距离徙动和多普勒徙动的现象,影响机动目标的积累与检测性能.针对上述问题,提出一种基于频率轴反转变换与广义变尺度傅里叶变换的机动目标检测快速非搜索算法.首先在距离频域-方位时域利用频率轴反转变换校正距离徙动,回波信号变为线性调频信号;接着利用Wigner-Ville分布变换核与广义变尺度傅里叶变换对目标参数进行估计;最后在距离-多普勒域完成目标能量的积累.与现有方法相比,所提方法可以快速校正距离徙动,实现非搜索的目标参数估计,计算复杂度低.仿真实验表明,该方法可有效完成机动目标的检测与参数估计.%In the detection for the maneuvering target,the integration and detection performance for the moving targets may slide down due to the defocusing effects of the range migration and the Doppler frequency migration caused by target radial velocity and acceleration.To address these issues,a fast non-searching method based on fast time-frequency axis reversal and generalized scaled Fourier transform is proposed.In this algorithm,the frequency axis reversal transform is firstly applied to correct the range migration.Then,the received signal becomes the linear frequency modulated (LFM) signal,and both Wigner-Ville distribution (WVD) function and generalized scaled Fourier transform are applied to estimate the acceleration and velocity of the moving target.Finally,the moving target is well focused in the range-Doppler pared with the existing methods,the presented algorithm can fast eliminate the range migration and obtain the motionparameters without any searching procedure and can achieve a good balance between the computational cost and the detection ability as well as parameters estimation performance.Several simulation experiments are provided to demonstrate the effectiveness of the proposed algorithm.【期刊名称】《系统工程与电子技术》【年(卷),期】2018(040)001【总页数】8页(P1-8)【关键词】机动目标检测;参数估计;频率轴反转变换;广义变尺度傅里叶变换【作者】贺雄鹏;廖桂生;许京伟;黄鹏辉;魏嘉琪【作者单位】西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071;西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071;西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071;西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071;西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071【正文语种】中文【中图分类】TN957.510 引言随着高性能武器装备的发展,以新一代战斗机(具有超声速和高机动等特点)为代表的运动目标,对传统雷达的探测能力提出了严峻的挑战,高机动目标检测与高精度运动参数估计研究受到广泛的关注[1-3]。

基于自适应对偶字典的磁共振图像的超分辨率重建

ＡｄａｐｔｉｖｅＤｕａｌＤｉｃｔｉｏｎａｒｙ
ＬＩＵＺｈｅｎ－ｑｉ，ＢＡ０Ｌｉ－ｊｕｎ，ＣＨＥＮＺｈｏｎｇ
ｒＤｅｐａｒｔｍｅｎｔｏｆＥｌｅｃｔｒｏｎｉｃＳｃｉｅｎｃｅ，ＸｉａｍｅｎＵｎｉｖｅｒｓｉｔｙ，Ｘｉａｍｅｎ３６１００５，Ｃｈｉｎａ）
刘振圻，包立君，陈忠
（厦门大学电子科学系，福建厦门３６１００５）
摘要：为了提高磁共振成像的图像质量，提出了一种基于自适应对偶字典的超分辨率去噪重建方法，在超分辨率重建过程中引入去噪功能，使得改善图像分辨率的同时能够有效地滤除图像中的噪声，实现了超分辨率重建和去噪技术的有机结合。该方法利用聚类一ＰｃＡ算法提取图像的主要特征来构造主特征字典，采用训练方法设计出表达图像细节信息的自学习字典，两者结合构成的自适应对偶字典具有良好的稀疏度和自适应性。实验表明，与其他超分辨率算法相比，该方法超分辨率重建效果显著，峰值信噪比和平均结构相似度均有所提高。
第２８卷第４期
２０１３年８月
பைடு நூலகம்光电技术应用
ＥＬＥＣＴＲＯ一０ＰＴＩＣＴＥＣＨＮＯＬＯＧＹＡＰＰＬＩＣＡＴ１０Ｎ
ＶＯ１．２８．ＮＯ．４
Ａｕｇｕｓｔ，２０１３
・
信号与信息处理・
基于自适应对偶字典的磁共振图像的超分辨率重建

以对称反对称分裂预条件处理GMRES（m）的不精确牛顿法潮流计算

第33卷第19期电网技术V ol. 33 No. 19 2009年11月Power System Technology Nov. 2009 文章编号：1000-3673（2009）19-0123-04 中图分类号：TM712 文献标志码：A 学科代码：470·40以对称反对称分裂预条件处理GMRES(m)的不精确牛顿法潮流计算刘凯1，陈红坤1，向铁元1，高志新2（1．武汉大学电气工程学院，湖北省武汉市 430072；2．中南电力设计院，湖北省武汉市 430072）Inexact Newton Flow Computation Based on Hermitian andSkew-Hermitian Splitting Preconditioners GMRES(m)LIU Kai1，CHEN Hong-kun1，XIANG Tie-yuan1，GAO Zhi-xin2（1．School of Electrical Engineering，Wuhan University，Wuhan 430072，Hubei Province，China；2．Central Southern Electric Power Design Institute，Wuhan 430072，Hubei Province，China）ABSTRACT: According to the feature that the correction equation of large-scale power grid is highly sparse, a method for inexact Newton power flow computation based on Hermitian and skew-Hermitian preconditioners is researched. By use of symmetric and skew-Hermitian splitting of matrix, a new type of preconditioner is proposed. Combining the new preconditioner with GMRES(m), both convergency and convergence rate of power flow computation can be improved. Power flow computation results of IEEE 300-bus system show that the proposed algorithm is effective.KEY WORDS: power flow calculation；Hermitian and skew-Hermitian splitting；generalized minimal residual algorithm (GMRES(m))；preconditioning摘要：针对大规模电力系统修正方程式高度稀疏的特点，研究了一种基于对称反对称预处理的不精确牛顿法。

基于小波去噪的自适应波束形成算法研究

基于小波去噪的自适应波束形成算法研究文小方;张玲华;高婉贞【摘要】语音增强算法是数字助听器的一项关键技术.广义旁瓣抵消器自适应波束形成算法对受固定干扰源影响的目标语音有良好的消噪效果.针对广义旁瓣抵消器自适应波束形成算法带来的语音泄漏以及残留噪声问题,在分析语音泄漏原因的基础上,引进了小波阈值去噪技术,提出了基于小波去噪的自适应波束形成算法.该算法在对原始语音进行小波阈值去噪后,将去噪后的语音信号通过GSC结构,从而得到输出语音信号.因小波阈值去噪既能提高语音通过GSC结构前的信噪比,又能使通过GSC结构后的语音更加准确,一定程度上减少了语音的泄漏,降低了残留噪声的发生,提高了语音的辨识度.实验结果表明,所提出的算法使得算法性能有所提高,从而达到了增强语音的目的.%Speech enhancement algorithm is a key technique in digital hearing aid.The Generalized Sidelobe Canceller (GSC) adaptive beamforming algorithm can effectively remove the fixed interference source noise in the target speech.Aimed at the residual noise and speech leakage problems caused by the generalized sidelobe canceller adaptive beamforming algorithm,on the basis of analysis on the reasons for speech leakage and introduction of wavelet threshold denoising technology,the adaptive beamforming algorithm based on wavelet denoising has been proposed.After the denoising of wavelet threshold value has employed to deal with original speech,the outputs of audio signal through GSC structure obtained.Since wavelet threshold,value denoising can improve the SNR of the speech signal effectively and can reduce speech leakage with certain content,and enhance the speech recognition.The experimentalresults show that performance of the improved algorithm and the effects of speech enhancement have been achieved.【期刊名称】《计算机技术与发展》【年(卷),期】2017(027)006【总页数】4页(P169-172)【关键词】广义旁瓣抵消器;语音泄漏;小波阈值去噪;语音辨识度【作者】文小方;张玲华;高婉贞【作者单位】南京邮电大学通信与信息工程学院,江苏南京210003;南京邮电大学通信与信息工程学院,江苏南京210003;南京邮电大学通信与信息工程学院,江苏南京210003【正文语种】中文【中图分类】TP301.6随着科技的不断发展,人类的生活环境变得复杂起来。

泛化迁移深度学习下的跨模态图像行人识别算法

第42卷　第1期吉林大学学报(信息科学版)Vol.42　No.12024年1月Journal of Jilin University (Information Science Edition)Jan.2024文章编号:1671⁃5896(2024)01⁃0137⁃06泛化迁移深度学习下的跨模态图像行人识别算法收稿日期:2022⁃10⁃13基金项目:西安明德理工学院科研基金资助项目(2021XY01L09)作者简介:蔡现龙(1976 ),男,陕西渭南人,西安明德理工学院讲师,主要从事计算机科学与技术研究,(Tel)86⁃189****7386(E⁃mail)2631069053@㊂蔡现龙,李　阳,陈　曦(西安明德理工学院信息工程学院,西安710124)摘要:针对由于受光照条件变化㊁行人身高差异等影响,致使监控视频图像在不同时刻的成像存在较大的跨模态差异问题,为准确识别跨模态图像中的行人,提出基于泛化迁移深度学习的跨模态图像行人识别算法㊂通过循环生成对抗网络(Cyele GAN:Cycle Generative Adversarial Network)形成跨模态图像,采用单目标图像处理对基准图分割处理,得到人体候选区域,在匹配图中搜索和其匹配的区域,得到人体区域的视差,通过视差提取人体区域的深度和透视特征㊂将注意力机制和跨模态行人识别相结合,分析两种不同类型图像的差异,将两个子空间映射到同一个特征空间,同时引入泛化迁移深度学习算法对损失函数度量学习,自动筛选跨模态图像的行人特征,最终通过模态融合模块将筛选的特征融合处理完成行人识别㊂实验结果表明,所提算法可以快速㊁准确地提取不同模态图像中的行人,识别效果较好㊂关键词:泛化迁移深度学习;跨模态图像;行人识别;特征提取中图分类号:TP311文献标志码:APedestrian Recognition Algorithm of Cross⁃Modal Image under Generalized Transfer Deep LearningCAI Xianlong,LI Yang,CHEN Xi(School of Information Engineering,Xi’an Mingde Institute of Technology,Xi’an 710124,China)Abstract :Due to the influence of changes in lighting conditions and pedestrian height differences,there are large cross modal differences in surveillance video images at different times.In order to accurately identify pedestrians in cross modal images,a pedestrian recognition algorithm based on generalized transfer depth learning is proposed.The cross modal image is formed through Cyele GAN(Cycle Generative Adversarial Network),and the reference map is segmented using single object image processing to obtain candidate human body regions.The matching regions are searched in the matching map to obtain the disparity of human body regions,and the depth and perspective features of human body regions are extracted through the disparity.The attention mechanism and cross modal pedestrian recognition are combined to analyze the differences between the two types of images.The two subspaces are mapped to the same feature space.And the generalized migration depth learning algorithm is introduced to learn the loss function measurement,automatically screen the pedestrian features of the cross modal images,and finally complete pedestrian recognition through the modal fusion module to fuse the filtered features.The experimental results show that the proposed algorithm can quickly and accurately extract pedestrians from different modal images,and the recognition effect is good.Key words :generalization transfer deep learning;cross⁃modal images;pedestrian recognition;feature extraction0　引　言由于在光照条件较差的环境中对单模态行人识别,无法满足相关领域对行人识别效果的预期要求,因此人们将深度学习技术应用于行人识别[1⁃2]中,并在对应的数据集中取得了较高的识别率㊂由于昼夜光照差异比较明显,导致跨模态的行人识别面临巨大挑战㊂目前人们针对跨模态行人识别方面的研究已有许多报道,如王留洋等[3]优先组建双模态特征提取网络,通过构建的网络对图像深度特征实行提取操作,增强处理全部特征后融合图像的全部像素信息,完成行人识别㊂Oh 等[4]利用多个图像区域(头部㊁身体等)的convnet 特征构建了行人识别框架,从时间和视点两方面分析了不同特征的重要性,利用人脸识别器实现了行人人脸识别㊂郑爱华等[5]采用双路模型提取不同模态下的全局特征,对其实行局部精细化处理,挖掘行人的结构化局部信息;通过标签和预测信息构建跨模态局部信息之间的关联,完成跨模态融合处理,确保各个特征之间相互补充,最终实现行人识别㊂为降低光照等因素引起的图像模态差异对行人识别效果的影响,笔者引入泛化迁移深度学习,提出一种跨模态图像行人识别算法㊂经实验测试结果表明,所提算法能有效降低行人识别时间,提升行人识别结果的准确性㊂1　跨模态图像行人识别模型设计1.1　跨模态图像行人特征提取由于受摄像机角度㊁外部环境等因素影响,使行人视频监控图像产生了较大的模态差异,为此需要将识别的行人视频设定为一个图像集,利用Cyele GAN 生成跨模态图像㊂由于人体的轮廓在图像集中近似为矩形,所以可借助矩形目标检测方法得到人体候选区域㊂优先采用Hough变换方法提取行人的主要图1　人体候选区域获取流程图Fig.1　Flow chart of human body candidate region acquisition 特征信息,通过视知觉分组的灰度分类器和共圆分类器将人体候选区域虚假信息剔除㊂图1给出了人体候选区域获取的详细操作流程图㊂为得到人体候选区域不同区域的特征信息,优先需要获取不同区域的视差㊂在实际操作过程中,采用基于局部约束的像素点区域匹配算法㊂以基准图中待匹配像素点为中心构建一个窗口,通过窗口内相邻像素的灰度值描述图像中的像素特征㊂将基准图中随机一个像素点设定为中心,同时创建多个大小完全一致的滑动窗口,引入搜索策略获取像点图在对准图中对应的像素点,两者之间的差值即为视差㊂块匹配方法[6⁃7]的核心是将基准图待匹配的窗口设定为模板图像,对准图像作为目标图像,对两者实行模板匹配㊂在匹配过程中,主要通过人体候选区域每个灰度间的相关测度描述不同视图间的相关性,如下:D p SSD (h )=∑(u ,w )∈R p R (u ,v )-I m R (u ,v ),(1)其中D p SSD (h )表示视图之间的相关性;R (u ,v )表示跨模态图像的水平偏移量;I m 表示基准图像;R p 表示随机像素对应的块状邻域㊂由于每个候选区域的相关性保持不变,所以需要将目标区域中区域相关性设定为式(1)的形式,进而获取目标区域对应的距离测度,如下:D T SAD (h )=∑(u ,w )∈R p 1R (u ,v )-I m R (u ,v ),(2)其中D T SAD (h )表示各个目标区域之间的距离测度㊂在实际应用过程中,需要消除左右两个视图之间由于光照亮度产生的差异,为此引入零均值方法,将其应用目标匹配过程中,进而获取零均值视图相关性D T ZSAD (h ),如下:831吉林大学学报(信息科学版)第42卷D T ZSAD (h )=∑(u ,w )∈R p R (u ,v )-I m -1R (u ,v )-I m R (u ,v )㊂(3) 通过候选人体区域取代式(2)和式(3)中的目标区域,而候选人体区域的视差可根据外极线约束在经过校正处理后的左右视图中,沿外极线方向搜索目标最小视图相关性D T ZSAD (h ),如下:[D T SAD (h )]min =arg min (u ,w )∈R p [D p SSD (h )-D T ZSAD (h )]㊂(4) 在跨模态图像中,人体和其他物体之间存在明显差异,则跨模态图像可能出现的行人身高最小值为h min ,如下:h min =H -b D T ZSAD (h ),(5)其中H 表示人体候选区域内的深度特征㊂设定人体区域在空间中的真实长度为l ,在采集人体图像的过程中,可通过小孔透视比例得到不同轮廓的特征提取结果:W (u ,v )=(z -h )h 1R (u ,v ),(6)其中z 表示人体候选区域的深度㊂由于跨模态图像中人体候选区域的视差半径和真实人体身高之间存在密切关联,而人体的真实身高可看做是行人的固有特征,设定行人身高的变化范围,则有h min ≤h ≤h max ㊂通过上述分析,利用图2给出跨模态图像行人特征提取流程图㊂图2　跨模态图像行人特征提取流程Fig.2　Flow chart of pedestrian feature extraction from cross⁃modal images 通过人体视觉[8⁃9]可得到人体区域的深度和透视特征,如下:S (u ,v )=1[D T SAD (h )]min R (u ,v )I m ,T (u ,v )={W (u ,v )(z -h )}2I m ìîíïïï,(7)其中S (u ,v )和T (u ,v )分别表示人体区域的深度特征和透视特征㊂1.2　泛化迁移深度学习下的跨模态图像行人识别深度学习中的注意力机制是指重点关注图像的细节信息,忽略没有利用价值的信息,使其在图像领域得到广泛应用,取得了十分显著的成果㊂将通道域思想应用于跨模态图像行人识别中,可以快速获取红绿蓝(RGB:Red,Green and Blue)和相对照度(RI:Relative Illumination)图像两者之间的差异性,进而准确区分不同类型的行人㊂通过SeNet 网络的思想全面引入压缩激活神经网络,其中压缩激活模块主要是利用每个通道之间的关系,学习特征权重,有效增强特征图关键信息的权重比例㊂设定输入特征为F ={f 1,f 2, ,f n },大小为F ∈E (h ,w ,c ),优先对1.1小节得到的特征压缩处理,通过全局池化的方式,将特征图转换为大小完全相同的向量,即全局通道描述符b (u ,v ),如下:b (u ,v )=F (sp )(u ,v ),1W (u ,v )∑m =1∑n =1f n (i ,j {),(8)931第1期蔡现龙,等:泛化迁移深度学习下的跨模态图像行人识别算法其中F (sp )(㊃)表示压缩操作;f n (i ,j )表示通道总数㊂通过两个全连接层得到特征向量u 的计算如下:u =H (u ,v )(i ,j ),β(g (u ,v {)),(9)其中H (u ,v )(㊃)表示激励操作;β表示激活函数;g (u ,v )表示两个全连接层对应的权值矩阵㊂将注意力机制应用于跨模态图像行人识别中,构建基于压缩激活机制的双路径跨模态模型,模型中融入了压缩激活模块,方便后续学习更加具有鲁棒性的特征㊂学习不同模态下的特征,将其映射到对应的子空间中㊂通过上述分析,优先计算行人各个特征之间的欧氏距离,并基于其再次计算即可获取三元组损失函数,如下:K chtri =1F (sp )(u ,v )∑m =1∑n =1f n (i ,j )[max(D (u ,v )-min D (u ,v ))+β],(10)其中K chtri 表示三元组损失函数;D (u ,v )表示相同跨模态图像之间的特征距离㊂将三元组损失函数和身份损失函数两者结合,最终获取综合损失函数如下:K tocal =K chtri +K id ,(11)其中K tocal 表示综合损失函数;K id 表示身份损失函数㊂经上述分析,引入泛化迁移深度学习算法对综合损失函数度量学习,则有:K tocal (u ,v )=(k a ,p -β)K chtri +K id ,(12)其中K tocal (u ,v )表示综合损失函数的度量学习结果;k a ,p 表示超参数㊂对输入的原始图像,通过测试集形成的跨模态图像集并没有得到充分应用,所以需要借助模态融合模块将两种筛选后的特征融合处理,同时将融合后的结果输入到全连接层中,采用SoftMax 损失展开有监督的训练㊂模态融合[10]模块的主要目的是将原始图像和跨模态图像两者有效融合,在设定条件下可利用RGB 图得到丰富的颜色特征,采用RI 图像可得到丰富的纹理特征,如下:L lsr =(1-β)lg{p (k )}-1/K chtri (k a ,p -β),(13)其中L lsr 表示跨模态图像的纹理特征;p (k )表示平滑参数㊂采用模态融合模块融合处理上述提取的特征和式(13)提取的纹理特征,以实现跨模态图像行人识别,如下:Q (u ,v )=1/(1-β){(k a ,p -β)K tocal (u ,v )}f n (i ,j ),(14)其中Q (u ,v )表示跨模态图像的行人识别结果㊂至此,实现跨模态图像行人识别㊂2　实验分析为验证所提泛化迁移深度学习下的跨模态图像行人识别算法的有效性,实验在INRIA Person Dataset 图像库(http:∥pascal.inrialpes.fr /data /human /)中随机选择200幅跨模态图像作为测试图像集,设定图像的大小为256×256像素,优先利用图3给出部分测试图像㊂图3　部分行人测试图像集Fig.3　Part of the pedestrian test image set 041吉林大学学报(信息科学版)第42卷将文献[3⁃4]算法作为所提方法的对比方法,从不同角度对图3所示的行人图像进行测试㊂2.1　实验流程实验计算机配备IntelXeon 6230(2.10GHz)CPU 和32GByte 视频内存的NVIDIA Tesla V100视频卡㊂实验中,文献[3⁃4]算法行人识别流程和参数设置依照其实验最佳参数进行设定㊂笔者算法具体的实验流程如图4所示㊂图4　所提算法识别流程Fig.4　Identification process of the proposed algorithm 2.2　实验结果分析在图3所示的测试图像集上进行实验测试,分析不同算法的识别效果,实验测试结果如图5所示㊂图5　不同算法的跨模态图像行人识别结果对比Fig.5　Comparison of pedestrian recognition results incross⁃modal images by different algorithms 从图5可看出,无论白天还是夜晚,采用所提算法均可准确识别行人,而另外两种算法在比较复杂的场景下只能识别出行人的局部特征信息,出现了漏识和误识现象㊂由此可见,所提算法利用模态融合模块能更好地完成行人识别,且受光照差异造成的模态差异影响较小㊂以相同数据集中不同光照强度的图像作为测试对象,将识别时间作为测试指标,表1给出了具体实验分析结果㊂表1　不同算法的跨模态图像行人识别时间测试结果对比　平均识别时间为1.732s,分别低于另外两种算法的1.79s 和1.85s,全面验证了笔者算法的优势,同时可以更快的速度完成行人识别,受光照影响较小㊂141第1期蔡现龙,等:泛化迁移深度学习下的跨模态图像行人识别算法图6　图像不同视差距离下峰值信噪比数值Fig.6　Peak signal to noise ratio values of images at different parallax distances 以峰值信噪比(PSNR:Peak Singal⁃Noise Ratio)为指标,测试在图像不同视差距离下行人识别的峰值信噪比数值,结果如图6所示㊂从图6可看出,随着视差距离的增大,行人识别图像峰值信噪比数值虽然呈现降低趋势,但降低幅度很小㊂其中笔者方法的峰值信噪比数值始终高于两种对比算法㊂上述结果说明笔者方法将泛化迁移深度学习引入到行人识别中,获取的行人识别结果较完整,表明识别能力较好㊂3　结　语针对行人识别方法受光照㊁视差距离影响产生的模态差异造成识别时间较长以及识别结果不准确的问题,笔者提出一种泛化迁移深度学习下的跨模态图像行人识别算法㊂通过和另外两种算法对比可知,笔者算法可以全面降低行人识别所用时间,同时还能增加识别结果准确性,为后续开展此方面研究提供了重要的策略和理论依据㊂参考文献:[1]祁磊,于沛泽,高阳.弱监督场景下的行人重识别研究综述[J].软件学报,2020,31(9):2883⁃2902.QI L,YU P Z,GAO Y.Research on Weak⁃Supervised Person Re⁃Identification [J].Journal of Software,2020,31(9):2883⁃2902.[2]韩光,葛亚鸣,张城玮.基于去相关高精度分类网络与重排序的行人再识别[J].计算机应用研究,2020,37(5):1587⁃1591,1596.HAN G,GE Y M,ZHANG C W.Person Re⁃Identification by Decorrelated High⁃Precision Classification Network and Re⁃Ranking [J].Application Research of Computers,2020,37(5):1587⁃1591,1596.[3]王留洋,芮挺,郑南,等.基于跨模态特征增强的RGB⁃T 行人检测算法研究[J].兵器装备工程学报,2022,43(5):254⁃260.WANG L Y,RUI T,ZHENG N,et al.Research on RGB⁃T Pedestrian Detection Algorithm Based on Cross⁃Modal Feature Enhancement [J].Journal of Ordnance Equipment Engineering,2022,43(5):254⁃260.[4]OH S J,BENENSON R,FRITZ M,et al.Person Recognition in Personal Photo Collections [J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2020,42(1):203⁃220.[5]郑爱华,曾小强,江波,等.基于局部异质协同双路网络的跨模态行人重识别[J].模式识别与人工智能,2020,33(10):867⁃878.ZHENG A H,ZENG X Q,JIANG B,et al.Cross⁃Modal Person Re⁃Identification Based on Local Heterogeneous CollaborativeDual⁃Path Network [J].Pattern Recognition and Artificial Intelligence,2020,33(10):867⁃878.[6]AGARWAL R,VERMA O P.Robust Copy⁃Move Forgery Detection Using Modified Superpixel Based FCM Clustering withEmperor Penguin Optimization and Block Feature Matching [J].Evolving Systems,2022,13(1):27⁃41.[7]JAVDANI D,RAHMANI H,WEISS G.SeMBlock:A Semantic⁃Aware Meta⁃Blocking Approach for Entity Resolution [J].Intelligent Decision Technologies:An International Journal,2021,15(3):461⁃468.[8]WU J Y,LU C H,LO H H,et al.P⁃23:Image Adaptation to Human Vision (Eyeglasses Free):Full Visual⁃CorrectedFunction in Light⁃Field Near⁃to⁃Eye Displays [J].SID International Symposium:Digest of Technology Papers,2021,52(3):1143⁃1145.[9]ANNAMALAI R,DORNEICH M,TOKADLI G.Evaluating the Effect of Poor Contrast Ratio in Simulated Sensor⁃Based VisionSystems on Performance [J].IEEE Transactions on Human⁃Machine Systems,2021,51(6):632⁃640.[10]邓佳桐,程志江,叶浩劼.改进YOLOv3的多模态融合行人检测算法[J].中国测试,2022,48(5):108⁃115.DENG J T,CHENG Z J,YE H J.Multimodal Fusion Pedestrian Detection Algorithm Based on Improved YOLOv3[J].China Measurement &Testing Technology,2022,48(5):108⁃115.(责任编辑:刘东亮)241吉林大学学报(信息科学版)第42卷。

纹理物体缺陷的视觉检测算法研究--优秀毕业论文

摘要
在竞争激烈的工业自动化生产过程中，机器视觉对产品质量的把关起着举足轻重的作用，机器视觉在缺陷检测技术方面的应用也逐渐普遍起来。与常规的检测技术相比，自动化的视觉检测系统更加经济、快捷、高效与安全。纹理物体在工业生产中广泛存在，像用于半导体装配和封装底板和发光二极管，现代化电子系统中的印制电路板，以及纺织行业中的布匹和织物等都可认为是含有纹理特征的物体。本论文主要致力于纹理物体的缺陷检测技术研究，为纹理物体的自动化检测提供高效而可靠的检测算法。纹理是描述图像内容的重要特征，纹理分析也已经被成功的应用与纹理分割和纹理分类当中。本研究提出了一种基于纹理分析技术和参考比较方式的缺陷检测算法。这种算法能容忍物体变形引起的图像配准误差，对纹理的影响也具有鲁棒性。本算法旨在为检测出的缺陷区域提供丰富而重要的物理意义，如缺陷区域的大小、形状、亮度对比度及空间分布等。同时，在参考图像可行的情况下，本算法可用于同质纹理物体和非同质纹理物体的检测，对非纹理物体的检测也可取得不错的效果。在整个检测过程中，我们采用了可调控金字塔的纹理分析和重构技术。与传统的小波纹理分析技术不同，我们在小波域中加入处理物体变形和纹理影响的容忍度控制算法，来实现容忍物体变形和对纹理影响鲁棒的目的。最后可调控金字塔的重构保证了缺陷区域物理意义恢复的准确性。实验阶段，我们检测了一系列具有实际应用价值的图像。实验结果表明本文提出的纹理物体缺陷检测算法具有高效性和易于实现性。关键字: 缺陷检测；纹理；物体变形；可调控金字塔；重构
Keywords: defect detection, texture, object distortion, steerable pyramid, reconstruction
II

CSI模型构建及其参数的GME的综合估计研究

CSI模型构建及其参数的GME的综合估计研究摘要本文主要讨论了CSI模型构建及其参数的GME综合估计方法。

CSI模型（common spatial patterns）是一种经典的模式识别算法，广泛应用于脑机接口、人体姿态识别等领域。

GME（generalized method of moments）是一种估计参数的方法，它使得最小化样本矩与理论矩之间的差异来估计参数。

本文将两种算法进行结合，于是得到CSI模型参数的GME估计方法，并且对于该方法进行分析和探讨。

首先，本文介绍了CSI模型的构建方法以及基本原理。

其次，在CSI模型的参数估计中，使用了GME方法。

GME方法简单易懂，且在较小的样本情况下仍然能够得到较为准确的结果。

而在实际应用中，CSI模型的参数较多，参数的估计往往会受到噪声的干扰，因此引入GME方法能够有效地抑制噪声的影响，得到更加准确的结果。

同时，本文还提出了一种正则化方法来控制参数的数量和大小，以及一个加速算法来提高算法的效率。

最后，本文通过实验进行了验证，证明了所提出方法的有效性和优越性。

实验结果表明，使用GME方法对CSI模型的参数进行估计，可以获得更加准确的结果，且可以有效地处理噪声的影响，实现了较高的准确度和稳定性。

关键词：CSI；GME；模式识别；参数估计IntroductionIn recent years, the CSI model has been widely used in pattern recognition and machine learning. It is an effective algorithm for feature extraction and dimensionality reduction, which has been applied to brain-computer interface, human posture recognition, and other fields. The CSI model uses thecommon spatial patterns (CSP) algorithm to maximize the differences between the classes of data after feature extraction. This improves the classification performance of the machine learning algorithm. However, the CSP algorithm needs to estimate a large number of parameters, which are influenced by noise. Therefore, it is necessary to develop a robust and efficient parameter estimation method for the CSI model.The generalized method of moments (GME) is a statistical method that can estimate parameters of models. The GME method minimizes the difference between empirical moments and theoretical moments. This method is widely used in econometrics, finance, and engineering. In recent years, it has been used in machine learning and pattern recognition. The GME method has advantages such as robustness, simplicity, and low computational cost. Therefore, we propose to use the GME method to estimate the parameters of the CSI model.Our research mainly focuses on the construction of the CSI model and the estimation of its parameters using the GME method. We propose a regularization method to control the number and size of the parameters. We also develop an accelerated algorithm to improve the efficiency of the algorithm. Finally, we validate the effectiveness and superiority of the proposed method through experiments.CSI model constructionThe CSI model is based on the CSP algorithm, which maximizes the differences between the classes of data. The CSP algorithm finds a set of spatial filters that project the EEG data into a new subspace where the differences between the classes are maximized. Let X be the EEG data, y be the class label, and c be the number of classes. Let S1 and S2 bethe covariance matrices of the EEG data for two classes. The CSP algorithm finds the projection matrix W that maximizes the following objective function:$$ W = \arg\max_{W} \text{tr}(W^T S_1 W) / \text{tr}(W^T S_2 W) $$The CSP algorithm decomposes the projection matrix Winto two sets of spatial filters: the left spatial filters L and the right spatial filters R. Let d be the number of channels. The left spatial filters L are a d by m matrix, where m is the number of filters. The right spatial filters R are a m by d matrix. The projection matrix W is obtained by multiplying L and R:$$ W = L^T R $$The left spatial filters L capture the most discriminative information, while the right spatial filters R remove the noise. Therefore, the left spatial filters are used to extract the features, and the right spatial filters are used to restore the original data.Parameter estimation using GMEThe CSP algorithm needs to estimate the covariance matrices S1 and S2 for two classes. The estimation of the covariance matrices is influenced by the noise in the EEG data. Therefore, we propose to use the GME method to estimate the parameters of the CSP algorithm. The GME method minimizes the difference between the empirical moments and the theoretical moments. Let θ be the parameter vector of the CSP algorithm. The empirical moments are calculated using the following formula:$$ \frac{1}{n} \sum_{i=1}^n g(y_i, x_i, \theta) $$where g is a function that depends on the parameter vector θ, x is the EEG data, and y is the class label. Thetheoretical moments are calculated using the following formula:$$ E[g(y,x,\theta)] = \int g(y,x,\theta) p(y,x) dx dy $$where p(y,x) is the joint probability distribution ofthe class label y and the EEG data x. By minimizing the difference between the empirical moments and the theoretical moments, we can obtain the parameter vector θ:$$ \hat{\theta} = \arg\min_{\theta} \frac{1}{n}\sum_{i=1}^n [g(y_i, x_i, \theta) - E[g(y,x,\theta)]]^2 $$ The GME method can be used to estimate the parameters of the CSP algorithm. However, the CSP algorithm has a large number of parameters, which are influenced by noise. Therefore, it is necessary to use a regularization method to control the number and size of the parameters.Regularization methodTo avoid overfitting and improve the generalization performance of the CSP algorithm, we propose to use a regularization method. Let λ be the regularization parameter, and let θ0 be the estimate of the parameter vector without regularization. The regularized estimate of the parameter vector is calculated using the following formula:$$ \hat{\theta}_{reg} = \arg\min_{\theta} \frac{1}{n}\sum_{i=1}^n [g(y_i, x_i, \theta) - E[g(y,x,\theta)]]^2 +\lambda ||\theta - \theta_0||^2 $$The regularization parameter λ controls the trade-off between the goodness of fit and the complexity of the model.A large value of λ leads to a simple model, while a small value of λ leads to a complex model. The regularization parameter can be selected using cross-validation.Accelerated algorithmThe GME method involves computing a large number ofintegrals, which are computationally expensive. Therefore, we propose an accelerated algorithm to improve the efficiency of the algorithm. The accelerated algorithm uses a fast Fourier transform (FFT) to compute the integrals. Let g be a function of the EEG data x, and let f(g) be the Fourier transform of g. The accelerated algorithm computes the theoretical moments using the following formula:$$ E[g(y,x,\theta)] \approx \sum_{\omega} f(g)(\omega)p(y,\omega|\theta) $$where p(y,ω|θ) is the Fourier transform of the joint probability distribution p(y,x|θ). The accelerated algorithm can reduce the computational complexity of the GME method.ExperimentWe conducted experiments on a publicly available dataset called BCI competition IV dataset IIa. The dataset contains EEG data of subjects with four motor imagery tasks: left hand, right hand, foot, and tongue. We used a 3 by 3 CSP algorithmto extract the features, and we used a linear discriminant analysis (LDA) algorithm to classify the features. We compared the performance of the following methods: the CSP algorithm with no regularization, the CSP algorithm with λ = 0.1, the CSP algorithm with λ = 1, and the CSP algorithmwith the accelerated algorithm.The results of the experiment are shown in Table 1. The CSP algorithm with λ = 1 achieved the best performance, with an accuracy of 75.8%. The CSP algorithm with the accelerated algorithm achieved an accuracy of 75.3%. The CSP algorithmwith no regularization achieved an accuracy of 73.7%. The CSP algorithm with λ = 0.1 achieved an accuracy of 73.2%. The results show that the proposed methods can achieve better classification performance than the CSP algorithm with noregularization. The regularization method and the accelerated algorithm can improve the accuracy and the efficiency of the algorithm.Table 1. Classification accuracy (%).Method AccuracyCSP (no regularization) 73.7CSP (λ = 0.1) 73.2CSP (λ = 1) 75.8CSP (accelerated) 75.3ConclusionIn this paper, we proposed a GME-based method to estimate the parameters of the CSP algorithm. The GME method can suppress the impact of noise and obtain more accurate parameter estimates. We also proposed a regularization method and an accelerated algorithm to improve the accuracy and the efficiency of the algorithm. The experimental results showed that the proposed methods can achieve better classification performance than the CSP algorithm without regularization. The proposed methods can be applied to real-worldapplications such as brain-computer interface and human posture recognition.。

MT-003(Understand SINAD,ENOB,SNR,THD,THD+N,and SFDR so you Don't Get Lost in the Noise Floor)

MT-003TUTORIAL Understand SINAD, ENOB, SNR, THD, THD + N, and SFDR soYou Don't Get Lost in the Noise Floorby Walt KesterINTRODUCTIONSix popular specifications for quantifying ADC dynamic performance are SINAD (signal-to-noise-and-distortion ratio), ENOB (effective number of bits), SNR (signal-to-noise ratio), THD (total harmonic distortion), THD + N (total harmonic distortion plus noise), and SFDR (spurious free dynamic range). Although most ADC manufacturers have adopted the same definitions for these specifications, some exceptions still exist. Because of their importance in comparing ADCs, it is important not only to understand exactly what is being specified, but the relationships between the specifications.There are a number of ways to quantify the distortion and noise of an ADC. All of them are based on an FFT analysis using a generalized test setup such as shown in Figure 1.Figure 1: Generalized Test Setup for FFT Analysis of ADC OutputThe spectral output of the FFT is a series of M/2 points in the frequency domain (M is the size of the FFT—the number of samples stored in the buffer memory). The spacing between the points is f s/M, and the total frequency range covered is dc to f s/2, where f s is the sampling rate. The width of each frequency "bin" (sometimes called the resolution of the FFT) is f s/M. Figure 2 shows an FFT output for an ideal 12-bit ADC using the Analog Devices' ADIsimADC® program. Note that the theoretical noise floor of the FFT is equal to the theoretical SNR plus the FFT process gain, 10×log(M/2). It is important to remember that the value for noise used in the SNR calculation is the noise that extends over the entire Nyquist bandwidth (dc to f s/2), but the FFT acts as a narrowband spectrum analyzer with a bandwidth of f s/M that sweeps over the spectrum. This has the effect of pushing the noise down by an amount equal to the process gain—the same effect as narrowing the bandwidth of an analog spectrum analyzer.The FFT data shown in Figure 2 represents the average of 5 individual FFTs. Note that averaging a number of FFTs does not affect the average noise floor, it only acts to "smooth" the random variations in the amplitudes contained in each frequency bin.Figure 2: FFT Output for an Ideal 12-Bit ADC,Input = 2.111MHz,f s = 82MSPS, Average of 5 FFTs, M = 8192,Data Generated from ADIsimADC®The FFT output can be used like an analog spectrum analyzer to measure the amplitude of the various harmonics and noise components of a digitized signal. The harmonics of the input signal can be distinguished from other distortion products by their location in the frequency spectrum. Figure 3 shows a 7-MHz input signal sampled at 20 MSPS and the location of the first 9 harmonics. Aliased harmonics of f a fall at frequencies equal to |±Kf s± nf a|, where n is the order of the harmonic, and K = 0, 1, 2, 3,.... The second and third harmonics are generally the only ones specified on a data sheet because they tend to be the largest, although some data sheets may specify the value of the worst harmonic.Harmonic distortion is normally specified in dBc (decibels below carrier), although in audio applications it may be specified as a percentage. It is the ratio of the rms signal to the rms value of the harmonic in question. Harmonic distortion is generally specified with an input signal near full-scale (generally 0.5 to 1 dB below full-scale to prevent clipping), but it can be specified at any level. For signals much lower than full-scale, other distortion products due to the differential nonlinearity (DNL) of the converter—not direct harmonics—may limit performance.FREQUENCY (MHz)Figure 3: Location of Distortion Products: InputSignal = 7 MHz, Sampling Rate = 20 MSPSTotal harmonic distortion (THD) is the ratio of the rms value of the fundamental signal to the mean value of the root-sum-square of its harmonics (generally, only the first 5 harmonics are significant). THD of an ADC is also generally specified with the input signal close to full-scale, although it can be specified at any level.Total harmonic distortion plus noise (THD + N) is the ratio of the rms value of the fundamental signal to the mean value of the root-sum-square of its harmonics plus all noise components (excluding dc). The bandwidth over which the noise is measured must be specified. In the case of an FFT, the bandwidth is dc to f s/2. (If the bandwidth of the measurement is dc to f s/2 (the Nyquist bandwidth), THD + N is equal to SINAD—see below). Be warned, however, that in audio applications the measurement bandwidth may not necessarily be the Nyquist bandwidth. Spurious free dynamic range (SFDR) is the ratio of the rms value of the signal to the rms value of the worst spurious signal regardless of where it falls in the frequency spectrum. The worst spur may or may not be a harmonic of the original signal. SFDR is an important specification in communications systems because it represents the smallest value of signal that can be distinguished from a large interfering signal (blocker). SFDR can be specified with respect to full-scale (dBFS) or with respect to the actual signal amplitude (dBc). The definition of SFDR is shown graphically in Figure 4.dBf sFREQUENCY2Figure 4: Spurious Free Dynamic Range (SFDR)The Analog Devices' ADIsimADC® ADC modeling program allows various high performance ADCs to be evaluated at varioius operating frequencies, levels, and sampling rates. The models yield an accurate representation of actual performance, and a typical FFT output for the AD9444 14-bit, 80-MSPS ADC is shown in Figure 5. Note that the input frequency is 95.111 MHz and is aliased back to 15.111 MHz by the sampling process. The output also displays the locations of the first five harmonics. In this case, all the harmonics are aliases. The program also calculates and tabulates the important performance parameters as shown in the left-hand data column.SINAD = 73.20dBSNR = 73.42dBTHD = 86.31dBSFDR = 89.03dBcNOISE FLOOR = 109.67dBM = 8192Figure 5: AD9444 14-Bit, 80MSPS ADC f in = 95.111MHz, f s = 80MSPS, Average of 5 FFTs, M = 8192, Data Generated from ADIsimADC®SIGNAL-TO-NOISE-AND-DISTORTION RATIO (SINAD), SIGNAL-TO-NOISE RATIO (SNR), AND EFFECTIVE NUMBER OF BITS (ENOB)SINAD and SNR deserve careful attention, because there is still some variation between ADC manufacturers as to their precise meaning. Signal-to-Noise-and-Distortion (SINAD, or S/(N + D) is the ratio of the rms signal amplitude to the mean value of the root-sum-square (rss) of all other spectral components, including harmonics, but excluding dc. SINAD is a good indication of the overall dynamic performance of an ADC because it includes all components which make up noise and distortion. SINAD is often plotted for various input amplitudes and frequencies. For a given input frequency and amplitude, SINAD is equal to THD + N, provided the bandwidth for the noise measurement is the same for both (the Nyquist bandwidth). A typical plot for the AD9226 12-bit, 65-MSPS ADC is shown in Figure 6.ANALOG INPUT FREQUENCY (MHz)Figure 6: AD9226 12-bit, 65-MSPS ADC SINAD and ENOBfor Various Input Full-Scale Spans (Range)The SINAD plot shows that the ac performance of the ADC degrades due to high-frequency distortion and is usually plotted for frequencies well above the Nyquist frequency so that performance in undersampling applications can be evaluated. SINAD plots such as these are very useful in evaluating the dynamic performance of ADCs. SINAD is often converted to effective-number-of-bits (ENOB) using the relationship for the theoretical SNR of an ideal N-bit ADC: SNR = 6.02N + 1.76 dB. The equation is solved for N, and the value of SINAD is substituted for SNR:02 .6dB 76.1SINAD ENOB −=. Eq.1Note that Equation 1 assumes a full-scale input signal. If the signal level is reduced, the value of SINAD decreases, and the ENOB decreases. It is necessary to add a correction factor for calculating ENOB at reduced signal amplitudes as shown in Equation 2:02.6Amplitude Input Amplitude Fullscale log 20db 76.1SINAD ENOB MEASURED ⎟⎟⎠⎞⎜⎜⎝⎛+−=. Eq. 2The correction factor essentially "normalizes" the ENOB value to full-scale regardless of the actual signal amplitude.Signal-to-noise ratio (SNR, or sometimes called SNR-without-harmonics ) is calculated from the FFT data the same as SINAD, except that the signal harmonics are excluded from the calculation, leaving only the noise terms. In practice, it is only necessary to exclude the first 5 harmonics, since they dominate. The SNR plot will degrade at high input frequencies, but generally not as rapidly as SINAD because of the exclusion of the harmonic terms.A few ADC data sheets somewhat loosely refer to SINAD as SNR, so you must be careful when interpreting these specifications and understand exactly what the manufacturer means.THE MATHEMATICAL RELATIONSHIPS BETWEEN SINAD, SNR, AND THDThere is a mathematical relationship between SINAD, SNR, and THD (assuming all are measured with the same input signal amplitude and frequency. In the following equations, SNR, THD, and SINAD are expressed in dB, and are derived from the actual numerical ratios S/N, S/D, and S/(N+D) as shown below:⎟⎠⎞⎜⎝⎛=N S log 20SNR , Eq. 3⎟⎠⎞⎜⎝⎛=D S log 20THD , Eq. 4⎟⎠⎞⎜⎝⎛+=D N S log 20SINAD .Eq. 5Eq. 3, Eq. 4, and Eq. 5 can be solved for the numerical ratios N/S, D/S, and (N+D)/S as follows:20/SNR 10SN−= Eq. 620/THD 10SD −= Eq. 720/SINAD 10SD N −=+Eq. 8Because the denominators of Eq. 6, Eq. 7, and Eq. 8 are all equal to S, the root sum square of N/S and D/S is equal to (N+D)/S as follows:21220/THD 220/SNR 21221010S D S N S D N ⎥⎦⎤⎢⎣⎡⎟⎠⎞⎜⎝⎛+⎟⎠⎞⎜⎝⎛=⎥⎥⎦⎤⎢⎢⎣⎡⎟⎠⎞⎜⎝⎛+⎟⎠⎞⎜⎝⎛=+−−, Eq. 92110/THD 10/SNR 1010SD N ⎥⎦⎤⎢⎣⎡+=+−−. Eq. 10 Therefore, S/(N+D) must equal:2110/THD 10/SNR 1010D N S −−−⎥⎦⎤⎢⎣⎡+=+, Eq. 11and hence,⎥⎦⎤⎢⎣⎡+−=⎟⎠⎞⎜⎝⎛+=−−10/THD 10/SNR 1010log 10D N S log 20SINAD . Eq. 12Eq. 12 gives us SINAD as a function of SNR and THD.Similarly, if we know SINAD and THD, we can solve for SNR as follows:⎥⎦⎤⎢⎣⎡−−=⎟⎠⎞⎜⎝⎛=−−10/THD 10/SINAD 1010log 10N S log 20SNR . Eq. 13Similarly, if we know SINAD and SNR, we can solve for THD as follows:⎥⎦⎤⎢⎣⎡−−=⎟⎠⎞⎜⎝⎛=−−10/SNR 10/SINAD 1010log 10D S log 20THD . Eq. 14Equations 12, 13, and 14 are implemented in an easy to use design tool on the Analog Devices' website. It is important to emphasize again that these relationships hold true only if the input frequency and amplitude are equal for all three measurements.SUMMARYBecause SINAD, SNR, ENOB, THD, THD + N, and SFDR are common measures of ADC dynamic performance, a complete understanding of them in the context of the manufacturers' data sheet is critical. This tutorial has defined the quantities and derived the mathematical relationship between SINAD, SNR, and THD.REFERENCES1.Walt Kester, Analog-Digital Conversion, Analog Devices, 2004, ISBN 0-916550-27-3,Chapter 6. Also available as The Data Conversion Handbook, Elsevier/Newnes, 2005, ISBN 0-7506-7841-0, Chapter 2.2.Hank Zumbahlen, Basic Linear Design, Analog Devices, 2006, ISBN: 0-915550-28-1.Also available as Linear Circuit Design Handbook, Elsevier-Newnes, 2008, ISBN-10:0750687037, ISBN-13: 978-0750687034. Chapter 6.Copyright 2009, Analog Devices, Inc. All rights reserved. Analog Devices assumes no responsibility for customer product design or the use or application of customers’ products or for any infringements of patents or rights of others which may result from Analog Devices assistance. All trademarks and logos are property of their respective holders. Information furnished by Analog Devices applications and development tools engineers is believed to be accurate and reliable, however no responsibility is assumed by Analog Devices regarding technical accuracy and topicality of the content provided in Analog Devices Tutorials.。

通感算一体化算法模型

通感算一体化算法模型The integration of synesthesia and algorithmic models poses several challenges in the field of artificial intelligence. Synesthesia is a neurological condition in which the stimulation of one sensory pathway leads to automatic, involuntary experiences in another sensory pathway. This phenomenon has been an area of interest for researchers in the development of algorithms that mimic the cross-modal sensory experiences of synesthetes. However, designing an algorithm that can accurately replicate such complex cognitive processes is no easy task. The intricate interplay between different sensory modalities and the subjective nature of synesthetic experiences present unique challenges for algorithm development.通感和算法模型的整合在人工智能领域面临着几个挑战。

通感是一种神经系统疾病，其中一个感觉通路的刺激会导致另一个感觉通路中自动、无意识的体验。

这种现象一直是研究人员感兴趣的领域，他们致力于开发模拟通感者交叉感官经验的算法。

数字信号处理英语词汇

AAbsolutely integrable绝对可积Absolutely integrable impulse response绝对可积冲激响应Absolutely summable绝对可和Absolutely summable impulse response绝对可和冲激响应Accumulator累加器Acoustic 声学Adder加法器Additivity property可加性Aliasing混叠现象All-pass systems全通系统AM (Amplitude modulation )幅度调制Amplifier放大器Amplitude modulation (AM)幅度调制Amplitude-scaling factor幅度放大因子Analog-to-digital (A-to-D) converter模数转换器Analysis equation分析公式（方程）Angel (phase) of complex number复数的角度（相位）Angle criterion角判据Angle modulation角度调制Anticausality反因果Aperiodic非周期Aperiodic convolution非周期卷积Aperiodic signal非周期信号Asynchronous异步的Audio systems音频（声音）系统Autocorrelation functions自相关函数Automobile suspension system汽车减震系统Averaging system平滑系统BBand-limited带（宽）限的Band-limited input signals带限输入信号Band-limited interpolation带限内插Bandpass filters带通滤波器Bandpass signal带通信号Bandpass-sampling techniques带通采样技术Bandwidth带宽Bartlett (triangular) window巴特利特（三角形）窗Bilateral Laplace transform双边拉普拉斯变换Bilinear双线性的Bilinear transformation双线性变换Bit（二进制）位，比特Block diagrams方框图Bode plots波特图Bounded有界限的Break frequency折转频率Butterworth filters巴特沃斯滤波器C“Chirp” transform algorithm“鸟声”变换算法Capacitor电容器Carrier载波Carrier frequency载波频率Carrier signal载波信号Cartesian (rectangular) form 直角坐标形式Cascade (series) interconnection串联，级联Cascade-form串联形式Causal LTI system因果的线性时不变系统Channel信道，频道Channel equalization信道均衡Chopper amplifier斩波器放大器Closed-loop闭环Closed-loop poles闭环极点Closed-loop system闭环系统Closed-loop system function闭环系统函数Coefficient multiplier系数乘法器Coefficients系数Communications systems通信系统Commutative property交换性（交换律）Compensation for nonideal elements非理想元件的补偿Complex conjugate复数共轭Complex exponential carrier复指数载波Complex exponential signals复指数信号Complex exponential(s)复指数Complex numbers 复数Conditionally stable systems条件稳定系统Conjugate symmetry共轭对称Conjugation property共轭性质Continuous-time delay连续时间延迟Continuous-time filter连续时间滤波器Continuous-time Fourier series连续时间傅立叶级数Continuous-time Fourier transform连续时间傅立叶变换Continuous-time signals连续时间信号Continuous-time systems连续时间系统Continuous-to-discrete-time conversion连续时间到离散时间转换Convergence 收敛Convolution卷积Convolution integral卷积积分Convolution property卷积性质Convolution sum卷积和Correlation function相关函数Critically damped systems临界阻尼系统Crosss-correlation functions互相关函数Cutoff frequencies截至频率DDamped sinusoids阻尼正弦振荡Damping ratio阻尼系数Dc offset直流偏移Dc sequence直流序列Deadbeat feedback systems临界阻尼反馈系统Decibels (dB) 分贝Decimation抽取Decimation and interpolation抽取和内插Degenerative (negative) feedback负反馈Delay延迟Delay time延迟时间Demodulation解调Difference equations差分方程Differencing property差分性质Differential equations微分方程Differentiating filters微分滤波器Differentiation property微分性质Differentiator微分器Digital-to-analog (D-to-A) converter数模转换器Direct Form I realization直接I型实现Direct form II realization直接II型实现Direct-form直接型Dirichlet conditions狄里赫利条件Dirichlet, P.L.狄里赫利Discontinuities间断点，不连续Discrete-time filters 离散时间滤波器Discrete-time Fourier series离散时间傅立叶级数Discrete-time Fourier series pair离散时间傅立叶级数对Discrete-time Fourier transform （DFT）离散时间傅立叶变换Discrete-time LTI filters离散时间线性时不变滤波器Discrete-time modulation离散时间调制Discrete-time nonrecursive filters离散时间非递归滤波器Discrete-time signals离散时间信号Discrete-time systems离散时间系统Discrete-time to continuous-time conversion离散时间到连续时间转换Dispersion弥撒（现象）Distortion扭曲，失真Distribution theory（property）分配律Dominant time constant主时间常数Double-sideband modulation (DSB)双边带调制Downsampling减采样Duality对偶性EEcho回波Eigenfunctions特征函数Eigenvalue特征值Elliptic filters椭圆滤波器Encirclement property围线性质End points终点Energy of signals信号的能量Energy-density spectrum能量密度谱Envelope detector包络检波器Envelope function包络函数Equalization均衡化Equalizer circuits均衡器电路Equation for closed-loop poles闭环极点方程Euler, L.欧拉Euler’s relation欧拉关系（公式）Even signals偶信号Exponential signals指数信号Exponentials指数FFast Fourier transform (FFT)快速傅立叶变换Feedback反馈Feedback interconnection反馈联结Feedback path反馈路径Filter(s)滤波器Final-value theorem终值定理Finite impulse response (FIR)有限长脉冲响应Finite impulse response (FIR) filters有限长脉冲响应滤波器Finite sum formula有限项和公式Finite-duration signals有限长信号First difference一阶差分First harmonic components基波分量（一次谐波分量）First-order continuous-time systems一阶连续时间系统First-order discrete-time systems一阶离散时间系统First-order recursive discrete-time filters一阶递归离散时间滤波器First-order systems一阶系统Forced response受迫响应Forward path正向通路Fourier series傅立叶级数Fourier transform傅立叶变换Fourier transform pairs傅立叶变换对Fourier, Jean Baptiste Joseph傅立叶（法国数学家，物理学家）Frequency response频率响应Frequency response of LTI systems线性时不变系统的频率响应Frequency scaling of continuous-time Fourier transform 连续时间傅立叶变化的频率尺度（变换性质）Frequency shift keying (FSK)频移键控Frequency shifting property频移性质Frequency-division multiplexing (FDM)频分多路复用Frequency-domain characterization频域特征Frequency-selective filter频率选择滤波器Frequency-shaping filters频率成型滤波器Fundamental components基波分量Fundamental frequency基波频率Fundamental period基波周期GGain增益Gain and phase margin增益和相位裕度General complex exponentials一般复指数信号Generalized functions广义函数Gibbs phenomenon吉伯斯现象Group delay群延迟HHalf-sample delay半采样间隔时延Hanning window汉宁窗Harmonic analyzer谐波分析议Harmonic components谐波分量Harmonically related谐波关系Heat propagation and diffusion热传播和扩散现象Higher order holds高阶保持Highpass filter高通滤波器Highpass-to-lowpass transformations高通到低通变换Hilbert transform希尔波特滤波器Homogeneity (scaling) property齐次性（比例性）IIdeal理想的Ideal bandstop characteristic理想带阻特征Ideal frequency-selective filter理想频率选择滤波器Idealization理想化Identity system恒等系统Imaginary part虚部Impulse response 冲激响应Impulse train冲激串Incrementally linear systems增量线性系统Independent variable独立变量Infinite impulse response (IIR)无限长脉冲响应Infinite impulse response (IIR) filters无限长脉冲响应滤波器Infinite sum formula无限项和公式Infinite taylor series无限项泰勒级数Initial-value theorem初值定理Inpulse-train sampling冲激串采样Instantaneous瞬时的Instantaneous frequency瞬时频率Integration in time-domain时域积分Integration property积分性质Integrator积分器Interconnection互联Intermediate-frequency (IF) stage中频级Intersymbol interference (ISI)码间干扰Inverse Fourier transform傅立叶反变换Inverse Laplace transform拉普拉斯反变换Inverse LTI system逆线性时不变系统Inverse system design逆系统设计Inverse z-transform z反变换Inverted pendulum倒立摆Invertibility of LTI systems线性时不变系统的可逆性Invertible systems逆系统LLag network滞后网络Lagrange, J.L.拉格朗日（法国数学家，力学家）Laplace transform拉普拉斯变换Laplace, P.S. de拉普拉斯（法国天文学家，数学家）lead network超前网络left-half plane左半平面left-sided signal左边信号Linear线性Linear constant-coefficient difference线性常系数差分方程equationsLinear constant-coefficient differential线性常系数微分方程equationsLinear feedback systems线性反馈系统Linear interpolation线性插值Linearity线性性Log magnitude-phase diagram对数幅－相图Log-magnitude plots对数模图Lossless coding无损失码Lowpass filters低通滤波器Lowpass-to-highpass transformation低通到高通的转换LTI system response线性时不变系统响应LTI systems analysis线性时不变系统分析MMagnitude and phase幅度和相位Matched filter匹配滤波器Measuring devices测量仪器Memory记忆Memoryless systems无记忆系统Modulating signal调制信号Modulation调制Modulation index调制指数Modulation property调制性质Moving-average filters移动平均滤波器Multiplexing多路技术Multiplication property相乘性质Multiplicities多样性NNarrowband窄带Narrowband frequency modulation窄带频率调制Natural frequency自然响应频率Natural response自然响应Negative (degenerative) feedback负反馈Nonanticipatibe system不超前系统Noncausal averaging system非因果平滑系统Nonideal非理想的Nonideal filters非理想滤波器Nonmalized functions归一化函数Nonrecursive非递归Nonrecursive filters非递归滤波器Nonrecursive linear constant-coefficient非递归线性常系数差分方程difference equationsNyquist frequency奈奎斯特频率Nyquist rate奈奎斯特率Nyquist stability criterion奈奎斯特稳定性判据OOdd harmonic 奇次谐波Odd signal奇信号Open-loop开环Open-loop frequency response开环频率响应Open-loop system开环系统Operational amplifier运算放大器Orthogonal functions正交函数Orthogonal signals正交信号Oscilloscope示波器Overdamped system过阻尼系统Oversampling过采样Overshoot超量PParallel interconnection并联Parallel-form block diagrams并联型框图Parity check奇偶校验检查Parseval’s relation帕斯伐尔关系（定理）Partial-fraction expansion部分分式展开Particular and homogeneous solution特解和齐次解Passband通频带Passband edge通带边缘Passband frequency通带频率Passband ripple通带起伏（或波纹）Pendulum钟摆Percent modulation调制百分数Periodic周期的Periodic complex exponentials周期复指数Periodic convolution周期卷积Periodic signals周期信号Periodic square wave周期方波Periodic square-wave modulating signal周期方波调制信号Periodic train of impulses周期冲激串Phase (angle) of complex number复数相位（角度）Phase lag相位滞后Phase lead相位超前Phase margin相位裕度Phase shift相移Phase-reversal相位倒置Phase modulation相位调制Plant工厂Polar form极坐标形式Poles极点Pole-zero plot(s)零极点图Polynomials 多项式Positive (regenerative) feedback正（再生）反馈Power of signals信号功率Power-series expansion method幂级数展开的方法Principal-phase function主值相位函数Proportional (P) control比例控制Proportional feedback system比例反馈系统Proportional-plus-derivative比例加积分Proportional-plus-derivative feedback比例加积分反馈Proportional-plus-integral-plus-different比例－积分－微分控制ial (PID) controlPulse-amplitude modulation脉冲幅度调制Pulse-code modulation脉冲编码调制Pulse-train carrier冲激串载波QQuadrature distortion正交失真Quadrature multiplexing正交多路复用Quality of circuit电路品质（因数）RRaised consine frequency response升余弦频率响应Rational frequency responses有理型频率响应Rational transform有理变换RC highpass filter RC 高阶滤波器RC lowpass filter RC 低阶滤波器Real实数Real exponential signals实指数信号Real part实部Rectangular (Cartesian) form 直角（卡笛儿）坐标形式Rectangular pulse矩形脉冲Rectangular pulse signal矩形脉冲信号Rectangular window矩形窗口Recursive (infinite impulse response)递归（无时限脉冲响应）滤波器filtersRecursive linear constant-coefficient 递归的线性常系数差分方程difference equationsRegenerative (positive) feedback再生（正）反馈Region of comvergence收敛域right-sided signal右边信号Rise time上升时间Root-locus analysis根轨迹分析（方法）Running sum动求和SS domain S域Sampled-data feedback systems采样数据反馈系统Sampled-data systems采样数据系统Sampling采样Sampling frequency采样频率Sampling function采样函数Sampling oscilloscope采样示波器Sampling period采样周期Sampling theorem采样定理Scaling (homogeneity) property比例性（齐次性）性质Scaling in z domain z域尺度变换Scrambler扰频器Second harmonic components二次谐波分量Second-order二阶Second-order continuous-time system二阶连续时间系统Second-order discrete-time system二阶离散时间系统Second-order systems二阶系统sequence序列Series (cascade) interconnection级联（串联）Sifting property筛选性质Sinc functions sinc函数Single-sideband单边带Single-sideband sinusoidal amplitude单边带正弦幅度调制modulationSingularity functions奇异函数Sinusoidal正弦（信号）Sinusoidal amplitude modulation正弦幅度调制Sinusoidal carrier正弦载波Sinusoidal frequency modulation正弦频率调制Sliding滑动Spectral coefficient频谱系数Spectrum频谱Speech scrambler语音加密器S-plane S平面Square wave方波Stability稳定性Stabilization of unstable systems不稳定系统的稳定性（度）Step response阶跃响应Step-invariant transformation阶跃响应不定的变换Stopband阻带Stopband edge阻带边缘Stopband frequency阻带频率Stopband ripple 阻带起伏（或波纹）Stroboscopic effect频闪响应Summer加法器Superposition integral叠加积分Superposition property叠加性质Superposition sum叠加和Suspension system减震系统Symmetric periodic 周期对称Symmetry对称性Synchronous同步的Synthesis equation综合方程System function(s)系统方程TTable of properties 性质列表Taylor series泰勒级数Time时间，时域Time advance property of unilateral单边z变换的时间超前性质z-transformTime constants时间常数Time delay property of unilateral单边z变换的时间延迟性质z-transformTime expansion property时间扩展性质Time invariance时间变量Time reversal property时间反转（反褶）性Time scaling property时间尺度变换性Time shifting property时移性质Time window时间窗口Time-division multiplexing (TDM)时分复用Time-domain时域Time-domain properties时域性质Tracking system (s)跟踪系统Transfer function转移函数transform pairs变换对Transformation变换（变形）Transition band过渡带Transmodulation (transmultiplexing) 交叉调制Triangular (Barlett) window三角型（巴特利特）窗口Trigonometric series三角级数Two-sided signal双边信号Type l feedback system l 型反馈系统UUint impulse response单位冲激响应Uint ramp function单位斜坡函数Undamped natural frequency无阻尼自然相应Undamped system无阻尼系统Underdamped systems欠阻尼系统Undersampling欠采样Unilateral单边的Unilateral Laplace transform单边拉普拉斯变换Unilateral z-transform单边z变换Unit circle单位圆Unit delay单位延迟Unit doublets单位冲激偶Unit impulse单位冲激Unit step functions单位阶跃函数Unit step response 单位阶跃响应Unstable systems不稳定系统Unwrapped phase展开的相位特性Upsampling增采样VVariable变量WWalsh functions沃尔什函数Wave波形Wavelengths波长Weighted average加权平均Wideband宽带Wideband frequency modulation宽带频率调制Windowing加窗zZ domain z域Zero force equalizer置零均衡器Zero-Input response零输入响应Zero-Order hold零阶保持Zeros of Laplace transform拉普拉斯变换的零点Zero-state response零状态响应z-transform z变换z-transform pairs z变换对。

Rsoft软件简介和使用教程

RSOFT使用教程目录Rsoft简介 (3)Chapter 7 Tutorials 第七章教程 (5)Tutorial 1: Ring Resonator 教程1：环形共振器 (5)Device Layout: 器件结构： (5)Defining Variables 定义变量 (6)Drawing the Structure 画器件结构图 (6)Checking the Index Profile 核对折射率分布 (9)Adding Time Monitors 添加时间监视（探测）器 (10)Simulation: Pulsed Excitation 模拟：脉冲激发 (12)Launch Field 激发场 (12)Wavelength/Frequency Spectrum 波长/频率光谱 (12)Increasing the Resolution of the FFT 提高FFT的分辨率 (14)Simulation: CW Excitation 模拟：连续激发 (16)Tutorial 2: PBG Crystal: Square Lattice 教程 2：PBG 晶体：四方晶格 (17)Lattice layout 晶格布局 (17)Base Lattice Generation 基准晶格的创建 (17)Lattice Customization 定制晶格 (18)Checking the Index Profile 核对折射率分布 (18)Inserting Time Monitors 插入时间监视器 (19)Launch Set Up 激发场设置 (20)Simulation 模拟 (21)Data Analysis 数据分析 (22)Switching Polarization 改变偏振为TM模 (23)Periodic Boundary Condition Set Up (24)Tutorial 3: PBG Crystal: Tee Structure 教程 3：PBG晶体： T型结构 (24)Tutorial 4: PBG Crystal: Defect Mode 教程四：PBG 晶体：缺陷模型 (24)Rsoft简介包括BeamPROP、FullWAVE、BandSOLVE、GratingMOD、DiffractMOD、FemSIM, 以及MOST软件。

Assembly optimised FFT algorithms for TMS320C62 processors

FFT dynamic range on fixed point DSPs
FFT dynamic range performance on fixed point DSPs heavily depends on the specific technique employed to prevent overflow. The simplest way to prevent overflow is feeding the FFT with a sufficiently small signal. For example, the TI FFT benchmark for ‘C62x DSP family has no procedure to control overflow. This makes the code very quick, but the input data corresponding to, e.g., a sinusoidal signal, must be limited to 9 significant bits, with a large dynamic range limitation. Fixed scaling represents a better way to avoid overflow. It involves a data division before every butterfly computation. The division factor is kept as low as possible 16 to limit the dynamic loss. For example, for the Radix-4 FFT, where a 3-bit scaling would by mathematically necessary, 2-bit scaling is typically preferred because it is sufficient in most practical cases. 16-bit The conditional scaling procedure examines the input data on every FFT stage alignment and carries out scaling only if necessary. In the block-floating-point FFT algorithm, data are aligned at a 16-bit format on every stage. In both cases, a dynamic improvement is obtained at the expense of longer computational times.

distance transform of sampled function解读

distance transform of sampled function解读Distance Transform of Sampled Function: An InterpretationIntroductionThe distance transform of a sampled function is a fundamental concept in digital image processing and computer vision. It serves as a powerful tool for various applications such as object recognition, image segmentation, and shape analysis. In this article, we will delve into the intricacies of the distance transform of a sampled function, its key properties, and its significance in computer science.Definition and Basic PrinciplesThe distance transform is an operation that assigns a distance value to each pixel in an image, based on its proximity to a specific target object or region. It quantifies the distance between each pixel and the nearest boundary of the object, providing valuable geometric information about the image.To compute the distance transform, first, a binary image is created, where the target object or region is represented by foreground pixels (usually white) and the background is represented by background pixels (usually black). This binary image serves as the input for the distance transform algorithm.Distance Transform AlgorithmsSeveral distance transform algorithms have been developed over the years. One of the most widely used algorithms is the chamfer distancetransform, also known as the 3-4-5 algorithm. This algorithm assigns a distance value to each foreground pixel by considering the neighboring pixels and their corresponding distances. Other popular algorithms include the Euclidean distance transform, the Manhattan distance transform, and the Voronoi distance transform.Properties of the Distance TransformThe distance transform possesses a set of important properties that make it a versatile tool for image analysis. These properties include:1. Distance Metric Preservation: The distance values assigned to the pixels accurately represent their geometric proximity to the boundary of the target object.2. Locality: The distance transform efficiently encodes local shape information. It provides a detailed description of the object's boundary and captures fine-grained details.3. Invariance to Object Shape: The distance transform is independent of the object's shape, making it robust to variations in object size, rotation, and orientation.Applications of the Distance TransformThe distance transform finds numerous applications across various domains. Some notable applications include:1. Image Segmentation: The distance transform can be used in conjunction with segmentation algorithms to accurately delineate objects inan image. It helps in distinguishing objects from the background and separating overlapping objects.2. Skeletonization: By considering the foreground pixels with a distance value of 1, the distance transform can be used to extract the object's skeleton. The skeleton represents the object's medial axis, aiding in shape analysis and recognition.3. Path Planning: The distance transform can assist in path planning algorithms by providing a distance map that guides the navigation of robots or autonomous vehicles. It helps in finding the shortest path between two points while avoiding obstacles.ConclusionThe distance transform of a sampled function plays a vital role in digital image processing and computer vision. Its ability to capture geometric information, preserve distance metrics, and provide valuable insights into the spatial structure of objects makes it indispensable in various applications. The proper understanding and utilization of the distance transform contribute to the advancement of image analysis techniques, enabling more accurate and efficient solutions in computer science.。

基于深度强化学习的无信号交叉口车辆协同控制算法

文章编号：1672-4747（2022）02-0014-11基于深度强化学习的无信号交叉口车辆协同控制算法蒋明智，吴天昊，张琳（北京邮电大学，人工智能学院，北京100876）摘要：针对未来智慧城市智能网联汽车通过无信号交叉口的通行效率问题，本文基于深度强化学习提出了一种渐进式价值期望估计的多智能体协同控制算法(PVE-MCC)。

设计了基于渐进式学习的价值期望估计策略，通过动态改变价值期望学习目标，保证值函数网络渐进式地持续学习，避免策略网络陷入局部最优解，并将该策略与泛化优势估计算法结合，提升算法收敛精度和稳定性。

其次，以通行效率、安全性和舒适性为优化目标，设计了多目标奖励函数来提高多智能体协同控制的综合性能。

此外，无信号交叉口易出现的“死锁”现象给多车协同控制带来了巨大的挑战，针对这一问题，基于链表环形检测算法设计了启发式的“死锁”检测-破解干预策略，实现对“死锁”环的提前检测和破解，进一步保障交通通行的安全性。

最后，本文搭建了双向六车道无信号交叉口场景的仿真实验平台，进行功能和性能验证。

实验结果表明，PVE-MCC 算法比现有方案提高交通流量30.47%，单车效率提升了95.56%，舒适性提升了53.82%。

关键词：智能交通；协同控制；强化学习；无信号交叉口；智能网联汽车中图分类号：U471.15文献标志码：ADOI ：10.19961/ki.1672-4747.2021.11.021Deep Reinforcement Learning Based Vehicular Cooperative ControlAlgorithm at Signal-free IntersectionJIANG Ming-zhi ，WU Tian-hao ，ZHANG Lin(School of Artificial Intelligence ，Beijing University of Posts and Telecommunications ，Beijing 100876，China)Abstract ：Aiming at the traffic efficiency of intelligent connected vehicles passing through a signal-free intersection in future smart cities ，in this paper we propose a progressive value-expectation esti-mation multi-agent cooperative control (PVE-MCC)algorithm based on deep reinforcement learn-ing.First ，the PVE-MCC algorithm designs a progressive value-expectation estimation (PVE)strate-gy based on progressive learning by dynamically varying the value expectation learning goal from short-term to long-term changes.The value function network is guaranteed to gradually and continu-ously learn ，and the strategic network is prevented from falling into a local optimal solution.Second ，the PVE-MCC algorithm combines the PVE strategy with the generalized advantage estimation algo-rithm to improve the convergence accuracy and stability of the algorithm.Third ，the PVE-MCC algo-rithm jointly takes traffic efficiency ，safety ，and comfort as the optimization objective ，and designs a multi-objective reward function to improve the performance of multi-agent collaborative control.In addition ，the “deadlock ”phenomenon that easily occurs at signal-free intersections constitutes a re-收稿日期：2021-11-16录用日期：2021-12-7网络首发：2021-12-17审稿日期：2021-11-16~11-21；11-26~12-02；12-06~12-07基金项目：产业技术基础公共服务平台项目(2019-00892-2-1)作者简介：蒋明智（1996—），男，硕士研究生，研究方向为智能交通系统，E-mail ：通信作者：张琳（1974—），男，教授，研究方向为移动云计算和物联网，E-mail ：引文格式：蒋明智，吴天昊，张琳.基于深度强化学习的无信号交叉口车辆协同控制算法[J].交通运输工程与信息学报，2022，20(2)：14-24.JIANG Ming-zhi ，WU Tian-hao ，ZHANG Lin.Deep Reinforcement Learning Based Vehicular Cooperative ControlAlgorithm at Signal-freeIntersection[J].Journal of Transportation Engineering and Information ，2022，20(2)：14-24.第20卷第2期2022年06月交通运输工程与信息学报Journal of Transportation Engineering and InformationVol.20No.2Jun.2022markable challenge for multi-vehicle cooperative control.In response to this problem，the PVE-MCC algorithm based on the linked list ring detection algorithm designs a heuristic detection-cracking in-tervention strategy for the“deadlock”to ensure the safety of the intersection.Finally，we present a simulation experimental platform for a two-way six-lane signal-free intersection for verification.The experimental results show that the PVE-MCC algorithm improves the traffic flow rate by30.47%，the single-vehicle efficiency by95.56%，and the comfort by53.82%compared with existing schemes.Key words：intelligent transportation；cooperative control；reinforcement learning；signal-free inter-section；intelligent connected vehicles0引言智慧城市的一个重要标志是城市交通智能化和汽车网联化，实现网联汽车在无信号交叉口高效智能地协同通行已成为当今国内外的研究热点。

ghzwidebandreal-timefftalgorithmbasedonfpga

GHz WIDEBAND REAL-TIME FFT ALGORITHMBASED ON FPGALuo Wenqiu1,a , Cao Peng1,b1School of Information Engineering, Beijing Institute of Graphic Communication,a******************.cn, b*******************.cnKeywords：Fast Fourier Transfer (FFT),Wideband, GHz, Field-Programmable Gate Array (FPGA), Parallel ComputingAbstractHigh-speed, wideband digital signal processing is required in the fields such as Radar, navigation, software radio, and so on. Improving the speed of FFT computation is one of the keys to realize high-speed real-time data acquisition and processing system. This paper proposes a solution to improve FFT computation by combining parallel computing approach on multi-processors and parallel computing architecture for FFT processor. The details of parallel processing algorithm and parallel computing architecture are given. ALtera FFT MegaCore is used to implement FFT, and computer simulation is given. The proposed method efficiently solves the problem of GHz wideband data processing.1 IntroductionRadar, navigation and software radio put forward higher request for high-speed data acquisition. With the rapid development of technology, the speed of ADC has increased greatly. High Speed ADC (>=1GSPS) products now are in available. The high speed ADC challenges the data processing ability of the real-time system. In order to improve the speed of FFT computation by the system, wideband and real-time processing are the key techniques[1].One method to improve the speed of FFT computation is parallel computing approach on multi-processor system. As described in literature [2], parallel computing on multiprocessor systems based on shared memory is used for FFT computation; in literature [3], parallel processing on multiple DSPs is used for large – point FFT computation. Another method is to advance the algorithms and structures of FFT processor. As described in literature [4], a two-dimensional decomposition algorithm is adopted to implement ultra high speed 32k point FFT processor; in literature [5], a fully parallel architecture is used to speed up FFT computation. As to the hardware implementation of FFT processor, there are three ways: general-purpose DSP processors approach, application specific integrated circuits (ASIC) approach and programmable logic devices (FPGA as representatives) approach. The features of DSP approach are short development time, high power consumption and low processing speed, so it is difficult to meet the high-speed, large-scale, real-time requirements. The features of ASIC approach are high speed FFT computation, low power consumption, but it is inflexible and difficult to expand. FPGA approach has the advantages of simple hardware structure, low power consumption and high flexibility[6-7].This paper researches on FFT algorithm to find a solution to increase FFT computation to meet the need of GHz, wideband data acquisition and processing.2Background of GHz wideband Data Acquisition and ProcessingIn high speed data acquisition system, when the speed of ACD is up to 1GSPS, it is hard for a single FFT processor to meet the real-time processing requirement. Aiming at this problem, parallel computing approach on multi-processors is adopted. Block diagram of parallel computing approach on multi-processors is shown in Fig.1. The system consists of ADC, ECL circuit and FFT processor based on FPGA. Where ADC implements analog-to-digital Convert, ECL circuit realizes the distribution and synthesis of high-speed signal, FPGA implements FFT computation.Fig.1 Block diagram of parallel computingapproach on multi-processors3 FFT AlgorithmThis paper presents two methods to improve FFT computation speed. One is parallel computing approach on multi-processor, and the other is to use parallel computing technology to implement high-radix algorithm. The aforementioned methods are described as follows.3.1FFT Algorithm Parallel DecompositionIn order to realize parallel computing on multi-processor, the output of ADC is transformed from one channel to multiple channel by ECL circuit. Data in each channel are processed by a single FFT processor. Taking 32-point FFT for example, data is represented as a matrix A which indicates that the data are assigned to K processors (in the following example, K take 4) in cyclic distribution, each row of matrix A represents the data input to a single processor. After 3)/(log 2=K N butterfly operation, these data is reassigned in block in these processors (that is,Transpose of the matrix A). Let matrix B represents transpose of the matrix A . Each processor sequentially stores 2/=K N rows of data, and completes four-point FFT computation for each row data. Where the cyclic distribution means the first K data are assigned sequentially to K processor, then the next K data are assigned to K processor, so on until all the data is done so far. Thus, the algorithm can be done in parallels.=3127231191511733026221814106229252117139512824201612840A=313029282726252423222120191817161514131211109876543210B3.2 Algorithm for Composite Parallel Computing ArchitectureCombining improved radix-4 algorithm and parallel / pipelined structure, this paper presents a composite parallel computing architecture to improve the speed of the FFT processor.Let )(n x denotes a length-N time-domain sequence, the discrete Fourier transform (DFT) is defined by∑−==1)()(N n nk N W n x k X ,10−≤≤N k （1）In order to complete the DFT computation in 4 parallel processing architecture, sequence x (n) is decomposed into 4 sub sequences, the discrete Fourier transform of each sub sequence is completed simultaneously in one clock cycle, 4 point data must be input to the 4 processing module in one clock cycle, So, Eq.(1) is transformed to Eq.(2):nNNn mn N nN N n mnN nNNn mn N Nn mn N W W N n jx N n x N n jx n x m X W W N n x N n x N n x n x m X W W N n jx N n x N n jx n x m X WN n x N n x N n x n x m X 31404/21404/1404/144/))4/3()2/()4/()((]34[))4/3()2/()4/()((]24[))4/3()2/()4/()((]14[))4/3()2/()4/()((]4[∑∑∑∑−=−=−=−=+−+−++=++−+++−=++++++−=+++++++=（2）Thus,∑∑∑∑−=−=−=−==+=+=+=1404/31404/21404/11404/0)(]34[,)(]24[)(]14[,)(]4[N n mnN N n mn N N n mn N N n mnN W n am X Wn am X Wn a m X W n am X （3）Where0a ,1a ,2a and 3a given bynNn NnNW N n jx N n x N n jx n x a WN n x N n x N n x n x a W N n jx N n x N n jx n x a N n x N n x N n x n x a 332210))4/3()2/()4/()(())4/3()2/()4/()(())4/3()2/()4/()(())4/3()2/()4/()((+−+−++=+−+++−=+++++−=++++++=So, a DFT computation can be completed by 4 independent 4/N point FFT module[8-9].4 SimulationsWith MATLAB / Simulink environment Altera DSP Builder can quickly complete fixed-point FFT processor design. In this paper, DSP Builder design flow is used to create a DSP Builder modelincluding a FFT MegaCore function variation. Altera's IP core FFT MegaCore is a high performance, parametric Fast Fourier Transform (FFT) processor IP core. The FFT MegaCore function can be parameterized to use either quad-output or single-output engine architecture. To increase the throughput of the FFT MegaCore function, quad-output is selected in this paper. Figure 2 is an illustration of quad -output enginearchitecture.Figure 2 quad -output engine architecture.Complex data samples X[k,0],X[k,1],X[k,2] andX[k,3] are read from internal memory in parallel and re-ordered by switch (SW). Next, the ordered samples are processed by the radix-4 butterfly processor to form the complex outputs. Because of the inherent mathematics of the radix-4 DIF decomposition, only three complex multipliers are requiredtoperformthethreenon-trivialtwiddle-factor multiplications on the outputs of the butterfly processor. In this architecture, thecalculation of all four 4 butterfly complex output are completed in one single clock cycle[10-12]. In this paper, the parameters of the FFT MegaCore function are set as follows:1024 point, 8 bits data precision, Streaming I/O data flow . The generated FFT processor module is shown in Fig.3. Input data loading sequence diagram is shown in Fig.4. Outputdata sequence diagramis shown in Fig.5.Figure 3 the generated FFT processor moduleFigure 4Input data loading sequence diagramFigure 5 Output data sequence diagramLet input signal x(t)=10sin(20πt), N=512, Sampling frequency is 500Hz. Simulation results of the sinusoidal signal based on IP core is shownin Fig.6. The result proves the method feasible.Figure 6 Simulation results5 ConclusionsIn order to meet the requirements of high-speed real-time signal processing, completing FFT computation quickly and flexibly is becoming increasingly important. This paper presents a solution to improve FFT computation by combining parallel computing approach on multi-processors and parallel computing architecture for FFT processor. It gives the details of parallel processing algorithm and parallel computing architecture. ALtera FFT Mega Core is used to implement FFT. It efficiently solves the problem of GHz high-speed data acquisition and processing.AcknowledgmentThis research is Supported by Beijing Natural Science Fund Project (4122025) and Beijing Natural Science Fund Key Project(B) (KZ2010100 15013).References[1]ZHENG Bing.: ‘Design and implementation of high-speed sampling buffer system based on FPGA’, Journal of Computer Applications, 2012, 32, (11) , pp. 3259–261[2] TIE Man-xia,DONG Yun-hong.: ‘Parallel Calculation of Fast Fourier Transform on Multi—Computer System’,Aeronautical Computing Technique, 2000,30, (3) , pp. 5–7．[3] LIULi,GAO Mei-Guo,ZhOU Run,WANGFei.: ‘Algorithm and Implementation of a Large-point FFT Under the Master-slave parallel Multi—Processor Architecture’, Journal of Systems Engineering and Electronics, 2000, 25, (10) , pp.1194–1196．[4]LI Wei, SUN Jinping, WANG Jun, LI Shaohong.: ‘ Implementation of 32 k points ultra high speed FFT processor based on FPGA devices’, Journal of Beijing University of Aeronautics and Astronautics. 2007,33, (12), pp. 1440–1443．[5]WANG Xudong, LIU Yu.: ‘ Full Parallel FFT Based on FPGA’, Journal of Nanjing University of Aeronautics and Astronautics, 2006, 38, (1) , pp. 96–100．[6]LI Jing-mei, ZHANG Chao-feng, JIN Yu-ping.: ‘Implementation of FFT Algorithm Based on FPGA’, Applied Science and Technology, 2009, 36,（2）, pp. 38–41．[7] HAN Ying, WANG Xu, WU Si Liang.: ‘The Study of High-Speed Windowed Complex FFT Processor Based on FPGA’, Journal of Beijing Institute of Technology, 2003,23, ( 3) , pp. 381–385.[8] FAN Jing,JIN Sheng-Zhen,SUN Cai-Hong.: ‘ Design and implementation of hyper-speed FFT Processor’, Optics and Precision Engineering, 2009, 17,（9）, pp. 2241–2245．[9]MA Zhuang,QI Lin,MA Pengge.: ‘ Implementation of FFT Based on FPGA IP Core’, Modern Electronics Technique, 2009,7, pp. 160–166．[10]DOU Xiu-mei,ZHAO Zhen-gang.: ‘Design and Implementation of FPGA FFT Algorithm Module Based on IP Core’, Radio Engineering, 2008, 38, (8) , pp. 29–31．[11]WU Wan-leng , S HAO Jie, X I A N Chu-hua.: ‘Efficient Sorting Architecture for Radix-4 FFT in FPGA’, Journal of Nanjing University of Aeronautics and Astronautic, 2005, 37, (2) , pp. 222–226．[12]LI Shizhuan, LI Weitao, JIANG Quanxian, FU Tian.: ‘A Fast FFT IP Core Design Based on Parallel Computing’, Computer & Digital Engineering, 2010, 38,（4）, pp. 139–141．。

稀疏恢复和傅里叶采样

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Leslie A. Kolodziejski Chair, Department Committee on Graduate Students
2
Sparse Recovery and Fourier Sampling by Eric Price
Submitted to the Department of Electrical Engineering and Computer Science on August 26, 2013, in partial fulﬁllment of the requirements for the degree of Doctor of Philosophy in Computer Science
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Department of Electrical Engineering and Computer Science August 26, 2013
Certiﬁed by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Piotr Indyk Professor Thesis Supervisor