Introduction Classification via Sparse Representation Conclusion Compressed Sensing Meets M
Robust Face Recognition via Sparse Representation
Robust Face Recognition via Sparse Representation -- A Q&A about the recent advances in face recognitionand how to protect your facial identityAllen Y. Yang (yang@)Department of EECS, UC BerkeleyJuly 21, 2008Q: What is this technique all about?A: The technique, called robust face recognition via sparse representation, provides a new solution to use computer program to classify human identity using frontal facial images, i.e., the well-known problem of face recognition.Face recognition has been one of the most extensively studied problems in the area of artificial intelligence and computer vision. Its applications include human-computer interaction, multimedia data compression, and security, to name a few. The significance of face recognition is also highlighted by a contrast between human’s high accuracy to recognize face images under various conditions and the computer’s historical poor accuracy.This technique proposes a highly accurate recognition framework. The extensive experiment has shown the method can achieve similar recognition accuracy as human vision, for the first time. In some cases, the method has outperformed what human vision can achieve in face recognition.Q: Who are the authors of this technique?A: The technique was developed in 2007 by Mr. John Wright, Dr. Allen Y. Yang, Dr. S. Shankar Sastry, and Dr. Yi Ma.The technique is jointly owned by the University of Illinois and the University of California, Berkeley. A provisional US patent has been filed in 2008. The technique is also being published in the IEEE Transactions on Pattern Analysis and Machine Intelligence [Wright 2008].Q: Why is face recognition difficult for computers?A: There are several issues that have historically hindered the improvement of face recognition in computer science.1.High dimensionality, namely, the data size is large for face images.When we take a picture of a face, the face image under certain color metrics will be stored as an image file on a computer, e.g., the image shown in Figure 1. Because the human brain is a massive parallel processor, it can quickly process a 2-D image and match the image with the other images learned in the past. However, the modern computer algorithms can only process 2-D images sequentially, meaning, it can only process an image pixel-by-pixel. Hence although the image file usually only takes less than 100 K Bytes to store on computer, if we treat each image as a sample point, it sits in a space of more than 10-100 K dimension (that is each pixel owns an individual dimension). Any pattern recognition problem in high-dimensional space (>100 D) is known to be difficult in the literature.Fig. 1. A frontal face image from the AR database [Martinez 1998]. The size of a JPEG file for this image is typically about 60 Kbytes.2.The number of identities to classify is high.To make the situation worse, an adult human being can learn to recognize thousands if not tens of thousands of different human faces over the span of his/her life. To ask a computer to match the similar ability, it has to first store tens of thousands of learned face images, which in the literature is called the training images. Then using whatever algorithm, the computer has to process the massive data and quickly identify a correct person using a new face image, which is called the test image.Fig. 2. An ensemble of 28 individuals in the Yale B database [Lee 2005]. A typical face recognition system needs to recognition 10-100 times more individuals. Arguably an adult can recognize thousands times more individuals in daily life.Combine the above two problems, we are solving a pattern recognition problem to carefully partition a high-dimensional data space into thousands of domains, each domain represents the possible appearance of an individual’s face images.3.Face recognition has to be performed under various real-world conditions.When you walk into a drug store to take a passport photo, you would usually be asked to pose a frontal, neutral expression in order to be qualified for a good passport photo. The store associate will also control the photo resolution, background, and lighting condition by using a uniform color screen and flash light. However in the real world, a computer program is asked to identify humans without all the above constraints. Although past solutions exist to achieve recognition under very limited relaxation of the constraints, to this day, none of the algorithms can answer all the possible challenges, including this technique we present.To further motivate the issue, human vision can accurately recognize learned human faces under different expressions, backgrounds, poses, and resolutions [Sinha 2006]. With professional training, humans can also identify face images with facial disguise. Figure 3 demonstrates this ability using images of Abraham Lincoln.Fig. 3. Images of Abraham Lincoln under various conditions (available online). Arguably humans can recognize the identity of Lincoln from each of these images.A natural question arises: Do we simply ask too much for a computer algorithm to achieve? For some applications such as at security check-points, we can mandate individuals to pose a frontal, neural face in order to be identified. However, in most other applications, this requirement is simply not practical. For example, we may want to search our photo albums to find all the images that contain our best friendsunder normal indoor/outdoor conditions, or we may need to identify a criminal suspect from a murky, low-resolution hidden camera who would naturally try to disguise his identity. Therefore, the study to recognize human faces under real-world conditions is motivated not only by pure scientific rigor, but also by urgent demands from practical applications.Q: What is the novelty of this technique? Why is the method related to sparse representation?A: The method is built on a novel pattern recognition framework, which relies on a scientific concept called sparse representation. In fact, sparse representation is not a new topic in many scientific areas. Particularly in human perception, scientists have discovered that accurate low-level and mid-level visual perceptions are a result of sparse representation of visual patterns using highly redundant visual neurons [Olshausen 1997, Serre 2006].Without diving into technical detail, let us consider an analogue. Assume that a normal individual, Tom, is very good at identifying different types of fruit juice such as orange juice, apple juice, lemon juice, and grape juice. Now he is asked to identify the ingredients of a fruit punch, which contains an unknown mixture of drinks. Tom discovers that when the ingredients of the punch are highly concentrated on a single type of juice (e.g., 95% orange juice), he will have no difficulty in identifying the dominant ingredient. On the other hand, when the punch is a largely even mixture of multiple drinks (e.g., 33% orange, 33% apple, and 33% grape), he has the most difficulty in identifying the individual ingredients. In this example, a fruit punch drink can be represented as a sum of the amounts of individual fruit drinks. We say such representation is sparse if the majority of the juice comes from a single fruit type. Conversely, we say the representation is not sparse. Clearly in this example, sparse representation leads to easier and more accurate recognition than nonsparse representation.The human brain turns out to be an excellent machine in calculation of sparse representation from biological sensors. In face recognition, when a new image is presented in front of the eyes, the visual cortex immediately calculates a representation of the face image based on all the prior face images it remembers from the past. However, such representation is believed to be only sparse in human visual cortex. For example, although Tom remembers thousands of individuals, when he is given a photo of his friend, Jerry, he will assert that the photo is an image of Jerry. His perception does not attempt to calculate the similarity of Jerry’s photo with all the images from other individuals. On the other hand, with the help of image-editing software such as Photoshop, an engineer now can seamlessly combine facial features from multiple individuals into a single new image. In this case, a typical human would assert that he/she cannot recognize the new image, rather than analytically calculating the percentage of similarities with multiple individuals (e.g., 33% Tom, 33% Jerry, 33% Tyke) [Sinha 2006].Q: What are the conditions that the technique applies to?A: Currently, the technique has been successfully demonstrated to classify frontal face images under different expressions, lighting conditions, resolutions, and severe facial disguise and image distortion. We believe it is one of the most comprehensive solutions in face recognition, and definitely one of the most accurate.Further study is required to establish a relation, if any, between sparse representation and face images with pose variations.Q: More technically, how does the algorithm estimate a sparse representation using face images? Why do the other methods fail in this respect?A: This technique has demonstrated the first solution in the literature to explicitly calculate sparse representation for the purpose of image-based pattern recognition. It is hard to say that the other extant methods have failed in this respect. Why? Simply because previously investigators did not realize the importance of sparse representation in human vision and computer vision for the purpose of classification. For example, a well-known solution to face recognition is called the nearest-neighbor method. It compares the similarity between a test image with all individual training images separately. Figure 4 shows an illustration of the similarity measurement. The nearest-neighbor method identifies the test image with a training image that is most similar to the test image. Hence the method is called the nearest neighbor. We can easily observe that the so-estimated representation is not sparse. This is because a single face image can be similar to multiple images in terms of its RGB pixel values. Therefore, an accurate classification based on this type of metrics is known to be difficult.Fig. 4. A similarity metric (the y-axis) between a test face image and about 1200 training images. The smaller the metric value, the more similar between two images. Our technique abandons the conventional wisdom to compare any similarity between the test image and individual training images or individual training classes. Rather, the algorithm attempts to calculate a representation of the input image w.r.t. all available training images as a whole. Furthermore, the method imposes one extra constraint that the optimal representation should use the smallest number of training images. Hence, the majority of the coefficients in the representation should be zero, and the representation is sparse (as shown in Figure 5).Fig. 5. An estimation of sparse representation w.r.t. a test image and about 1200 training images. The dominant coefficients in the representation correspond to the training images with the same identity as the input image. In this example, the recognition is based on downgraded 12-by-10 low-resolution images. Yet, the algorithm can correctly identify the input image as Subject 1.Q: How does the technique handle severe facial disguise in the image?A: Facial disguise and image distortion pose one of the biggest challenges that affect the accuracy of face recognition. The types of distortion that can be applied to face images are manifold. Figure 6 shows some of the examples.Fig. 6. Examples of image distortion on face images. Some of the cases are beyond human’s ability to perform reliable recognition.One of the notable advantages about the sparse representation framework is that the problem of image compensation on distortion combined with face recognition can be rigorously reformulated under the same framework. In this case, a distorted face image presents two types of sparsity: one representing the location of the distorted pixels in the image; and the other representing the identity of the subject as before. Our technique has been shown to be able to handle and eliminate all the above image distortion in Figure 6 while maintaining high accuracy. In the following, we present an example to illustrate a simplified solution for one type of distortion. For more detail, please refer to our paper [Wright 2008].Figure 7 demonstrates the process of an algorithm to recognize a face image with severe facial disguise by sunglasses. The algorithm first partitions the left test image into eight local regions, and individually recovers a sparse representation per region. Notice that with the sunglasses occluding the eye regions, the corresponding representations from these regions do not provide correct classification. However, when we look at the overall classification result over all regions, the nonocclused regions provide a high consensus for the image to be classified as Subject 1 (as shownin red circles in the figure). Therefore, the algorithm simultaneously recovers the subject identity and the facial regions that are being disguised.Fig. 7. Solving for part-based sparse representation using local face regions. Left: Test image. Right: Estimation of sparse representation and the corresponding classification on the titles. The red circle identifies the correct classiciation.Q: What is the quantitative performance of this technique?A: Most of the representative results from our extensive experiment have been documented in our paper [Wright 2008]. The experiment was based on two established face recognition databases, namely, the Extended Yale B database [Lee 2005] and the AR database [Martinez 1998].In the following, we highlight some of the notable results. On the Extended Yale B database, the algorithm achieved 92.1% accuracy using 12-by-10 resolution images, 93.7% using single-eye-region images, and 98.3% using mouth-region images. On the AR database, the algorithm achieves 97.5% accuracy on face images with sunglasses disguise, and 93.5% with scarf disguise.Q: Does the estimation of sparse representation cost more computation and time compared to other methods?A: The complexity and speed of an algorithm are important to the extent that they do not hinder the application of the algorithm to real-world problems. Our technique uses some of the best-studied numerical routines in the literature, namely, L-1 minimization to be specific. The routines belong to a family of optimization algorithms called convex optimization, which have been known to be extremely efficient to solve on computer. In addition, considering the rapid growth of the technology in producing advanced micro processors today, we do not believe there is any significant risk to implement a real-time commercial system based on this technique.Q: With this type of highly accurate face recognition algorithm available, is it becoming more and more difficult to protect biometric information and personal privacy in urban environments and on the Internet?A: Believe it or not, a government agency, a company, or even a total stranger can capture and permanently log your biometric identity, including your facial identity, much easier than you can imagine. Based on a Time magazine report [Grose 2008], a resident living or working in London will likely be captured on camera 300 times per day! One can believe other people living in other western metropolitan cities are enjoying similar “free services.” If you like to stay indoor and blog on the Internet, your public photo albums can be easily accessed over the nonprotected websites, and probably have been permanently logged by search engines such as Google and Yahoo!.With the ubiquitous camera technologies today, completely preventing your facial identity from being obtained by others is difficult, unless you would never step into a downtown area in big cities and never apply for a driver’s license. However, there are ways to prevent illegal and involuntary access to your facial identity, especially on the Internet. One simple step that everyone can choose to do to stop a third party exploring your face images online is to prevent these images from being linked to your identity. Any classification system needs a set of training images to study the possible appearance of your face. If you like to put your personal photos on your public website and frequently give away the names of the people in the photos, over time a search engine will be able to link the identities of the people with the face images in those photos. Therefore, to prevent an unauthorized party to “crawl” into your website and sip through the valuable private information, you should make these photo websites under password protection. Do not make a large amount of personal images available online without consent and at the same time provide the names of the people on the same website.Previously we have mentioned many notable applications that involve face recognition. The technology, if properly utilized, can also revolutionize the IT industry to better protect personal privacy. For example, an assembly factory can install a network of cameras to improve the safety of the assembly line but at the same time blur out the facial images of the workers from the surveillance videos. A cellphone user who is doing teleconferencing can activate a face recognition function to only track his/her facial movements and exclude other people in the background from being transmitted to the other party. All in all, face recognition is a rigorous scientific study. Its sole purpose is to hypothesize, model, and reproduce the image-based recognition process with accuracy comparable or even superior to human perception. The scope of its final extension and impact to our society will rest on the shoulder of the government, the industry, and each of the individual end users. References[Grose 2008] T. Grose. When surveillance cameras talk. Time (online), Feb. 11, 2008.[Lee 2005] K. Lee et al.. Acquiring linear subspaces for face recognition under variable lighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 5, 2005.[Martinez 1998] A. Martinez and R. Benavente. The AR face database. CVC Tech Report No. 24, 1998.[Olshausen 1997] B. Olshausen and D. Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision Research, vol. 37, 1997.[Serre 2006] T. Serre. Learning a dictionary of shape-components in visual cortex: Comparison with neurons, humans and machines. PhD dissertation, MIT, 2006.[Sinha 2006] P. Sinha et al.. Face recognition by humans: Nineteen results all computer vision researchers should know about. Proceedings of the IEEE, vol. 94, no. 11, November 2006.[Wright 2008] J. Wright et al.. Robust face recognition via sparse representation. (in press) IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008.。
基于CST仿真软件的阻抗匹配设计教学实验
ISSN 1002-4956 CN11-2034/T实验技术与管理Experimental Technology and Management第38卷第2期2021年2月Vol.38 No.2 Feb. 2021D O I: 10.16791/ki.sjg.2021.02.044基于C ST仿真软件的阻抗匹配设计教学实验赓臻\賡志斌2,刘宇平2(1.杭州电子科技大学电子信息学院,浙江杭州310018;2.新余学院数学与计算机学院,江西新余338000 )摘要:传输线的阻抗匹配是电磁场与微波技术中一个重要的理论,是射频微波电路设计的基础:但相关概念较为抽象,传统教学过程以数学推导为主,学生理解困难。
为了增强学生对阻抗匹配的理解,以微带线阻抗匹配的典型工程应用为案例,将理论分析与电磁仿真相结合,对微带线阻抗匹配网络进行设计,增强学生对传输线阻抗匹配的理解:使学生从理论到仿真,从数学推导到可视化的验证,构建全面的知识体系,增强 解决复杂工程问题的能力。
关键词:阻抗匹配;单支节匹配网络;微带线;电磁仿真中图分类号:G433 文献标识码:A 文章编号:1002-4956(2021)02-0204-04Teaching experiment of impedance matching designbased on CST simulation softwareLIAO Zhen1,LIAO Zhibin2,LIU Yuping2(1. School of Electronics and Information, Hangzhou Dianzi University, Hangzhou 310018, China;2. School of Mathematics and Computer, Xinyu University, Xinyu 338000, China)A bstract: The theory o f transmission line impedance matching is an important theory in electromagnetic field and microwave technology, and it is the fundamental o f radio and microwave circuit design. But the relative concepts are abstract and teaching process is based on mathematical derivation, which makes it difficult for students to understand. By taking a typical project o f the microstrip impedance matching as an example, the impedance matching network is designed by combining theoretical deduction with simulation, which has enhanced students’understanding o f transmission line impedance. The experiment is helpful to construct a comprehensive knowledge structure from theory to simulation and from formula deprivation to visual presentation and enhance students1 ability to solve complex engineering problems.Key w ords: impedance matching; single-stub matching network; microstrip; electromagnetic simulation随着通信技术的蓬勃发展,社会对射频微波技术 人才的需求也与日俱增+3]。
基于稀疏表达的人脸遮挡物去除
基于稀疏表达的人脸遮挡物去除吴从中;刘渠芬;詹曙【摘要】Face recognition technology is one of the most promising biometric technologies .Glasses , scarves and other obstructions have great impact on face recognition .In order to improve the recogni‐tion rate of the occluded face images ,a new occlusion removal method based on sparse representation is presented .In this method ,the sparse coefficients for the human face images with occlusion in the training sets are obtained .Then the images are reconstructed with the obtained coefficients to get the unoccluded face images .The experimental results show that this method can effectively remove the frontal occlusion and improve the rate of face recognition .%人脸识别技术是目前最具发展潜力的生物特征识别技术之一。
眼镜、围巾等遮挡物的存在对人脸识别系统的识别率影响很大,为了提高有遮挡的正面人脸图像的识别率,文章提出了基于稀疏表达分类的去除遮挡的方法。
该方法对于有遮挡的人脸图像先求出其在无遮挡人脸图像训练集上的稀疏系数,再根据求得的稀疏系数进行恢复重建,得到去遮挡的人脸图像。
天津大学微电子学与固体电子学考研复习辅导资料及导师分数线信息
天津大学微电子学与固体电子学考研复习辅导资料及导师分数线信息天津大学微电子学与固体电子学考研科目包括政治、外语、数学一以及电路、半导体物理或电介质物理、信号与系统、物理化学。
主要研究方向有四个,每个方向不同考试科目也不同,考生备考时需分清。
专业代码、名称及研究方向考试科目备注080903微电子学与固体电子学①101思想政治理论②201英语一或202俄语或203日语③301数学一④811电路①101思想政治理论②201英语一或202俄语或203日语③301数学一④813半导体物理或电介质物理①101思想政治理论②201英语一或202俄语或203日语③301数学一④815信号与系统①101思想政治理论②201英语一或202俄语或203日语③301数学一④839物理化学天津大学微电子学与固体电子学近两年考研录取信息院(系、所) 专业 报考人数 录取人数电子信息与工程学院(2012年)微电子学与固体电子学184 53电子信息与工程微电子学与固体193 52学院(2013年)电子学天津大学微电子学与固体电子学2012年报考人数为184人,录取人数为53人,2013年报考人数为193人,录取人数为52人。
由真题可以发现,现在考点涉及的广度和深度不断扩宽和加深。
由天津考研网签约的天津大学在读本硕博团队搜集整理了天津大学电子信息与工程学院微电子学与固体电子学考研全套复习资料,帮助考生梳理知识点并构建知识框架。
真题解析部分将真题按照知识点划分,条理清晰的呈现在同学们眼前。
然后根据各个考点的近几年真题解析,让同学对热点、难点了然于胸。
只有做到了对真题规律和趋势的把握,8—10月底的提高复习才能有的放矢、事半功倍!天津大学电子信息与工程学院微电子学与固体电子学考研导师信息刘开华纵向课题经费课题名称情境感知服务位置信息获取机理与算法 2009-01-01--2011-12-31 负责人:刘开华科技计划:国家基金委拨款单位:国家基金委合同经费:32课题名称智能航空铅封技术研究 2010-01-01--2012-12-31 负责人:刘开华科技计划:天津市科技支撑计划重点项目拨款单位:天津市科学技术委员会合同经费:50 横向课题经费课题名称基于相位法的RFID定位技术 2013-01-01--2013-12-31 负责人:刘开华科技计划: 拨款单位:中兴通信有限公司合同经费:16课题名称基于ADoc芯片组的产品开发 2008-09-01--2009-08-31 负责人:刘开华科技计划: 拨款单位:THOMSON宽带研发(北京)有限公司合同经费:6.3 期刊、会议论文Tan, Lingling; Bai, Yu; Teng, Jianfu; Liu, Kaihua; Meng, Wenqing Trans-Impedance Filter Synthesis Based on Nodal Admittance Matrix Expansion CIRCUITS SYSTEMS AND SIGNAL PROCESSINGnullTan, Lingling; Liu, Kaihua; Bai, Yu; Teng, Jianfu Construction of CDBA and CDTA behavioral models and the applications in symbolic circuits analysis ANALOG INTEGRATED CIRCUITS AND SIGNAL PROCESSINGnullMa Yongtao,Zhou Liuji,Liu Kaihua A Subcarrier-Pair Based Resource Allocation Scheme Using SensorsnullMa Yongtao,Zhou Liuji,Liu Kaihua, Wang Jinlong Iterative Phase Reconstruction and Weighted IEEE sensorsnull罗蓬,刘开华,闫格基于FrFT能量重心谱校正的LFM信号参数估计信号处理null 潘勇, 刘开华,等 A novel printed microstrip antenna with frequency reconfigurable characteristics for Bluetooth/WLAN/WiMAX applications Microwave and Optical Technology Lettersnull阎格,刘开华,吕西午基于分数阶Fourier变换的新型时频滤波器设计哈尔滨工业大学学报nullLin Zhu, Kaihua Liu, Zhang Qijun, Yongtao Ma and Bo Peng An enhanced analyticalNeuro-Space Mapping method for large-signal microwave device modeling null 罗蓬,刘开华,于洁潇,马永涛一种相干宽带线性调频信号的波达方向估计新方法通信学报nullLin Zhu, Yongtao Ma, Qijun Zhang and Kaihua Liu An enhanced Neuro-Space Mapping method for nonlinear device modeling nullYue Cui, Kaihua Liu, Junfeng Wang Direction-of-arrival estimation for coherent GPS signals based on oblique projection Signal ProcessingnullLV Xi-wu, LIU Kai-hua, et al. Efficient solution of additional base stations in time-of-arrival positioning systems Electronics Lettersnull省部级以上获奖刘开华;等数字电视接收系统、软件技术的研发与应用”天津市科技进步奖三等奖 2011-04-29李华;刘开华;等数字视频压缩与码流测试技术的研发及应用天津市科技进步奖二等奖 2009-04-29知识产权刘开华, 于洁潇高速公路上车辆的车速和相对位置实时测量系统及方法刘开华;潘勇;于洁潇;陈征一种基于无联网的车载自动实时监控远程终端刘开华,黄翔东,于洁潇,王兆华,闫格基于相位差测距的RFID无线定位方法王安国纵向课题经费课题名称基带处理与天线协同 2007-07-16--2011-11-16 负责人:王安国科技计划:国家科技部拨款单位:财政部合同经费:157.41课题名称无线网络多源稀疏协作编码研究 2011-01-01--2013-12-31 负责人:韩昌彩科技计划:国家基金委拨款单位:国家基金委合同经费:20横向课题经费课题名称具有波束多选择性的多频段可重构天线研究 2013-01-01--2014-12-31 负责人:王安国科技计划: 拨款单位:东南大学毫米波国家重点实验室合同经费:5 课题名称双方向图算法在室内定位中的应用 2012-01-01--2012-12-31 负责人:冷文科技计划: 拨款单位:中兴通讯股份有限公司合同经费:14.5期刊、会议论文马宁王安国姬雨初石和平Cooperative Space Shift Keying for Multiple-Relay Network IEEE Communications Lettersnull裴静王安国高顺,冷文Miniaturized Triple-Band Antenna With a Defected Ground Plane for WLAN/WiMAX Applications IEEE Antennas and Wireless Propagation Lettersnull 赵国煌王安国冷文陈彬陈华Wideband internal antenna with coupled feeding for 4G mobile phone Microwave and Optical Technology Lettersnull陈彬王安国赵国煌Design of a novel ultrawideband antenna with dual band-notchedcharacteristics Microwave and Optical technology lettersnull蔡晓涛王安国马宁冷文 A Novel Planar Parasitic Array Antenna with Reconfigurable Azimuth pattern IEEE Antennas and Wireless Propagation Lettersnull马宁王安国聂仲尔曲倩倩姬雨初Adaptive Mapping Generalized Space Shift Keying Modulation China Communicationsnull王安国蔡晓涛冷文带寄生贴片的圆盘形方向图可重构天线设计电波科学学报null 王安国陈彬冷文赵国煌一种小型化五频段可重构蝶形天线的设计电波科学学报null蔡晓涛王安国马宁冷文 Novel radiation pattern reconfigurable antenna with six beam choices The Journal of China Universities of Posts and Telecommunicationsnull 曲倩倩王安国聂仲尔郑剑锋 Block Mapping Spatial Modulation Scheme for MIMO Systems The Journal of China Universities of Posts and Telecommunicationsnull 王安国刘楠兰航方向图可重构宽带准八木天线的设计天津大学学报null李锵纵向课题经费课题名称基于稀疏核支持向量机的音乐自动分类系统关键技术研究2009-06-01--2010-06-01 负责人:李锵科技计划: 拨款单位:天津大学建筑设计研究院合同经费:3课题名称jg预研项目 2010-03-01--2010-12-01 负责人:李锵科技计划: 拨款单位:渤海石油运输有限责任公司合同经费:3课题名称超声波热治疗中非侵入式温度成像与弹性成像关键技术研究2015-01-01--2018-12-31 负责人:李锵科技计划:国家自然科学基金项目拨款单位: 国家自然科学基金委员会合同经费:85课题名称高等学校学科创新引智计划综合管理平台的设计与开发2010-04-01--2012-04-01 负责人:李锵科技计划: 拨款单位:苏州国芯科技有限公司合同经费:3横向课题经费课题名称微粒捕集器数据采集系统开发 2008-01-01--2008-06-01 负责人:李锵科技计划: 拨款单位:润英联新加坡私人有限公司合同经费:22.5课题名称电子系统可靠性增长建模与仿真 2006-12-01--2008-01-01 负责人:李锵科技计划: 拨款单位:中国人民解放军海军航空工程学院合同经费:5期刊、会议论文李锵,滕建辅,赵全明,李士心Wavelet domain Wiener filter and its application in signal denoising null张立毅,李锵,刘婷,滕建辅The research of the adaptive blind equalizer's steady residual error null徐星,李锵,关欣Chinese folk instruments classification via statistical features and sparse-based representation null张立毅,李锵,刘婷,滕建辅Study of improved constant modulus blind equalization algorithm null张立毅,孙云山,李锵,滕建辅Study on the fuzzy neural network classifier blind equalization algorithm null郭继昌,滕建辅,李锵Research of the gyro signal de-noising method based on stationary wavelets transform null肖志涛,于明,李锵,国澄明Symmetry phase congruency: Feature detector consistent with human visual system characteristics nullCai wei,李锵,关欣 Automatic singer identification based on auditory features. null 李锵,滕建辅,王昕,张雅绮,郭继昌Research of gyro signal de-noising with stationary wavelets transform null郭继昌,滕建辅,李锵,张雅绮The de-noising of gyro signals by bi-orthogonal wavelet transform nullLiu Tianlong,李锵,关欣Double boundary periodic extension DNA coding sequence detection algorithm combining base content null关欣,滕建辅,李锵,苏育挺Blind acoustic source separation combining time-delayed autocorrelation and 4TH-order cumulants null张立毅,李锵,滕建辅Kurtosis-driven variable step size blind equalization algorithm with constant module nullQin Lu,李锵,关欣Pitch Extraction for Musical Signals with Modified AMDF nullZhang Xueying,李锵,关欣 The Improved AMDF Gene Exon Prediction null李锵,Jian Dong,Ming-Guo Wang,滕建辅 Analysis and simulation of antenna protocol optimization for ad hoc networks nullFeng Yanyan,李锵,关欣Entropy of Teager Energy in Wavelet-domain Algorithm Applied in Note Onset Detection nullBao Hu, Li ShangSheng, 李锵,滕建辅Research on the technology of RFSS in large-scale universal missile ATE null张立毅,Haiqing Cheng,李锵,滕建辅 A research of forward neural network blind equalization algorithm based on momentum term null张立毅,李锵,滕建辅 A New Adaptive Variable Step-size Blind Equalization Algorithm Based on Forward Neural Network nullYutao Ma,李锵,Chao Li,Kun Li,滕建辅Design of active transimpedance band-pass filters with different Q values International Journal of Electronicsnull夏静静,李锵,刘浩澧,Wen-shiang Chen,Po-Hsiang Tsui An Approach for the Visualization of Temperature Distribution in Tissues According to Changes in Ultrasonic Backscattered Computational and Mathematical Methods in Medicinenull耿晓楠,李锵,崔博翔,王荞茵,刘浩澧超声温度影像与弹性成像监控组织射频消融南方医科大学学报null谭玲玲, 李锵, 李瑞杰, 滕建辅 Design of transimpedance low-pass filters InternationalJournal of Electronicsnull李锵,李秋颖,关欣基于听觉图像的音乐流派自动分类天津大学学报(自然科学与工程技术版)nullChong Zhou, Wei Pang, 李锵, Hongyu Yu, Xiaotang Hu, HaoZhang, Extracting the Electromechanical Coupling Constant of Piezoelectric Thin Film by the High-Tone Bulk Acoustic Resonator IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Controlnull 朱琳, 李锵, 刘开华基于ADS的声表面波单端对谐振器建模压电与声光null董丽梦, 李锵, 关欣基于稀疏表示分类器的音乐和弦识别系统研究计算机工程与应用null关欣,李锵,田洪伟基于差分全相位MFCC的音符起点自动检测计算机工程null 关欣,李锵,郭继昌,滕建辅二、四阶组合时延统计量多乐器盲分离计算机工程与应用null杨甲沛, 李锵, 刘郑, 袁晓琳基于自适应学习速率的改进型BP算法研究计算机工程与应用null李锵, 张法朝, 张瑞峰System design of DPF data recorder and data analysis null李锵, 袁晓琳, 杨甲沛Application of ant colony algorithm in the optimization of the time environmental conversion factor of the reliability models null张立毅,白煜,李锵,滕建辅复数系统中五二阶归一化积累盲均衡算法的研究通信学报null郭继昌,关欣,李锵,刘志杨红外图像预处理系统中模拟视频输出时序设计电子技术应用null关欣,滕建辅,李锵,苏育挺,Wang Shu-Yan Blind source separation combiningtime-delayed second and fourth order statistics 天津大学学报(自然科学与工程技术版)null 张立毅,李锵,滕建辅复数系统中三、二阶归一化累积量盲均衡算法的研究计算机工程与应用null张立毅,李锵,滕建辅经典盲均衡算法中稳态剩余误差的分析天津大学学报null 滕建辅,董健,李锵,关欣Design of maximally flat FIR filters based on explicit formulas combined with optimization 天津大学学报(英文版)null郭继昌,陈敏俊,李锵,关欣红外焦平面失效元处理方法及软硬件实现光电工程null 马杰,王昕,李锵,滕建辅基于特征值和奇异值分解方法的盲分离天津大学学报(自然科学与工程技术版)null李锵,郭继昌,关欣,滕建辅基于通用DSP的红外焦平面视频图像数字预处理系统天津大学学报(自然科学与工程技术版)null李锵,郭继昌,关欣,刘航,童央群基于DSP的红外焦平面视频图像数字处理系统的设计测控技术null马杰,滕建辅,李锵具有参考噪声源的多路传感器信号盲分离方法测控技术null 周郭飞,李锵,滕建辅微带扇形分支线在低通滤波器设计中应用电子测量技术null 李锵,滕建辅,李士心,肖志涛小波域Wiener滤波器信号的去噪方法天津大学学报(自然科学与工程技术版)null肖志涛,于明,李锵,唐红梅,国澄明 Log Gabor小波性能分析及其在相位一致性中应用天津大学学报(自然科学与工程技术版)null罗批,李锵,郭继昌,滕建辅Improved genetic algorithm and its performance analysis 天津大学学报(英文版)null罗批,郭继昌,李锵,滕建辅一种实用的电子线路参数优化算法电路与系统学报null 罗批,李锵,郭继昌,滕建辅基于偏最小二乘回归建模的探讨天津大学学报null 知识产权李锵,闫志勇,关欣一种结合SVM和增强型PCP特征的和弦识别方法中国2014100089231李锵, 冯亚楠, 关欣基于Teager能量熵的音符切分方法学术专著(关欣, 杨爱萍, 白煜, 李锵), 信号检测与估计:理论与应用(译著), 电子工业出版社2012-01-31(白煜, 李锵), 模拟集成电路设计的艺术(译著), 人民邮电出版社 2010-11-04(李锵,周进等), 无线通信基础(译著), 人民邮电出版社 2007-06-30(李锵,董健,关欣,鲍虎), 数字通信(原书第2版)(译著), 机械工业出版社 2006-02-28 (张为,关欣,刘艳艳,李锵), 电子电路设计基础(译著), 电子工业出版社 2005-10-01 (张雅绮,李锵等), Verilog HDL高级数字设计(译著), 电子工业出版社 2005-01-31 (李锵,侯春萍,赵宇), 网络(原书第2版)(译著), 机械工业出版社 2004-11-30(李锵,郭继昌), 无线通信与网络, 电子工业出版社 2004-06-30本文内容摘自《天津大学814通信原理考研红宝书》,更多考研资料可登陆网站下载!。
频域下稀疏表示的大数据库人脸分类算法
频域下稀疏表示的大数据库人脸分类算法胡业刚;任新悦;李培培;王汇源【摘要】人脸识别的识别率受众多因素影响,目前已有很多成形的高识别率算法,然而,随着数据库中人脸图像的增加,识别率下降很快。
鉴于该特点,采用频域下的稀疏表示分类算法能有效解决上述问题,先使用快速傅里叶变换(FFT)将人脸数据从时域变换到频域,再通过 l 1范数最优化稀疏表示算法,把所有训练样本作为基向量,稀疏表示出测试样本,最后使用最近邻子空间算法分类。
在扩展的 YaleB 人脸库中实验结果表明,该算法具有有效性。
%The recognition rate of face recognition is influenced by many factors, in which there are lots of effective algo-rithms, however, with the increase of face in the database, and the recognition rate will be decreased rapidly. In this situation, the sparse representation classification under the frequency domain can solve the above problems effectively. Firstly, the face image will be transformed from time domain to frequency domain using FFT algorithm, and then sparse representation about the test sample will be obtained by l1 norm optimization approach, in which all the training samples as the base vectors, in addition using the nearest neighbor subspace classification. Finally the experimental results show that the algorithm is effective in the extensional Yale B face database.【期刊名称】《阜阳师范学院学报(自然科学版)》【年(卷),期】2015(000)002【总页数】4页(P83-86)【关键词】稀疏表示;快速傅里叶变换;人脸识别【作者】胡业刚;任新悦;李培培;王汇源【作者单位】阜阳师范学院数学与统计学院,安徽阜阳 236037;阜阳师范学院数学与统计学院,安徽阜阳 236037;阜阳师范学院数学与统计学院,安徽阜阳236037;阜阳师范学院数学与统计学院,安徽阜阳 236037【正文语种】中文【中图分类】TP391.41 引言近年来,人脸识别已成为经典的模式识别研究问题之一。
基于压缩感知的鲁棒性人脸表情识别
基于压缩感知的鲁棒性人脸表情识别施徐敢;张石清;赵小明【摘要】为了有效提高噪声背景下的人脸表情识别性能,提出一种基于压缩感知的鲁棒性人脸表情识别方法。
先通过对腐蚀的测试样本表情图像进行稀疏表示,再利用压缩感知理论寻求其最稀疏的解,然后采用求得的最稀疏解信息实现人脸表情的分类。
在标准的Cohn-Kanade表情数据库的实验测试结果表明,该方法取得的人脸表情识别性能优于最近邻法、支持向量机以及最近邻子空间法。
可见,该方法用于人脸表情识别,识别效果较好,鲁棒性较高。
%In order to effectively improve the performance of facial expression recognition under the noisy background, a method of robust facial expression recognition based on compressed sensing is proposed. Firstly, the sparse representation of corrupted expression images of the identified test sample is sought, then the compressed sensing theory is used to solve its sparsest solution. Finally, according to the sparsest solution, facial expression classification is performed. Experimental results on benchmarking Cohn-Kanade database show that facial expression performance obtained by this method is better than the nearest neighbor (NN), support vector machine (SVM) and the nearest subspace (NS). Therefore, the proposed method shows both good recognition performance and high robustness on facial expression recognition tasks.【期刊名称】《计算机系统应用》【年(卷),期】2015(000)002【总页数】4页(P159-162)【关键词】压缩感知;稀疏表示;表情识别;鲁棒性;腐蚀【作者】施徐敢;张石清;赵小明【作者单位】浙江理工大学机械自动控制学院,杭州 310018; 台州学院图像处理与模式识别研究所,临海 317000;台州学院图像处理与模式识别研究所,临海317000;浙江理工大学机械自动控制学院,杭州 310018; 台州学院图像处理与模式识别研究所,临海 317000【正文语种】中文人脸表情是人们观察情感的重要标志, 如何使得机器能够认识人脸表情, 是一个既实用又有趣的研究方向. 如何让机器自动、高效、准确地来识别人类的情绪状态, 比如高兴、悲伤、愤怒、恐惧等, 即所谓的“人脸表情识别”[1]方面的研究, 是当前信号处理、模式识别、计算机视觉等领域的热点研究课题. 该研究在智能人机交互、人工智能等方面有着重要的应用价值.尽管人脸表情识别经过了多年的发展, 已经取得了较多的研究成果, 但现有的人脸表情识别研究[2-10]大多没有考虑表情图像受到噪声的影响. 在自然环境中, 人脸表情图像的获取、传输和存贮过程中常常也会受到各种噪声(如姿态、光照、腐蚀、遮挡等)的干扰而使图像降质, 从而导致人脸表情识别的性能会随之下降. 因此, 如何提高人脸表情识别的鲁棒性仍然是一个亟需解决的问题.压缩感知(Compressed sensing)或压缩采样(Compressive sampling)[11,12], 是近年来新出现的一种信号采样理论, 它可以在远小于Nyquist采样率的条件下获得信号的离散样本, 然后通过非线性重建无失真的完美信号. 压缩感知理论指出, 采样速率由信号中的内容和结构所决定, 而不再决定于信号的带宽. 目前, 压缩感知理论在图像处理[13]、人脸识别[14]、视频追踪[15]等领域受到了研究者的高度关注, 并表现出了极其强大的生命力, 但在人脸表情识别领域, 尤其针对鲁棒性的人脸表情识别问题, 国内外相关的文献报道甚少.压缩感知理论研究的初衷主要用于信号的压缩和表示, 但其最稀疏的表示具有很好的判别性. 本文利用压缩感知理论中的稀疏表示分类(Sparse Representation-based Classification, SRC)思想[14], 提出一种基于压缩感知的鲁棒性人脸表情识别方法. 先通过对腐蚀的测试样本表情图像进行稀疏表示, 再利用压缩感知理论寻求其最稀疏的解, 然后采用求得的最稀疏解信息实现人脸表情的分类. 在标准的Cohn-Kanade表情数据库[16]的实验结果表明了该方法的可行性.设A=[A1,A2,…,AC]是一组训练样本集, 总数量为n, 其中为第i类训练样本,y∈Rm是第i类的测试样本, 它可以由线性表示为:然而在实际情况中, 由于测试样本的类别一般是未知的, 所以式1可以写为式中, .由矩阵原来可知, m>n时, 矩阵(2)有唯一解; 但是在大多数情况下, m≤n, 此时矩阵(2)有无穷多个解. 为了使测试样本能够用自身所在类的训练样本进行线性表示, 这样的话系数向量x0中的非零向量应该尽可能少些. 所以对矩阵(2)求解可转换对矩阵(3)进行求解式中, ||·||0 表示l0范数, 它的作用是计算向量中非零元素的个数. 但是, 式(3)的求解非常困难, 这是个NP难题.由压缩感知理论可知: 当所求的系数足够稀疏时,可以把最小化l1范数的NP难题转化成最小化l1范数问题来求解.因此, 把式(3)改写为:然而在实际情况中, 获得的数据中经常含有噪声, 因此y很难由A进行比较准确的线性表示, 因此, 把式(4)改写为式(5)可以通过以下的式(6)来求解SRC算法可归纳如下:1)对训练样本集A中的每一个列向量进行归一化.2)求解最小化l1范数问题:或者求解3) 计算残差4) . 是的标记.本文采用标准的Cohn-Kanade[16]数据库进行实验. 通过对Cohn-Kanade数据库的原始图像采样得到32×32像素图像, 然后分别采用稀疏表示分类方法SRC、最近邻法(Nearest neighbor, NN), 支持向量机(Support Vector Machine, SVM), 以及近年来流行的最近邻子空间法(Nearest subspace, NS)[17]进行人脸表情识别实验, 并比较它们的性能.除了SRC方法, 使用的其它分类方法的基本思想表述如下: 最近邻法(NN)是基于样本学习的K近邻分类器(KNN), 当K=1时的一种情况. 支持向量机(SVM)是一种基于统计学习理论的分类器. 本文SVM采用“一对一”多类分类算法, 核函数为径向基函数, 并对核函数参数值进行最优化, 即在训练样本数据上使用交叉验证方法实现. 最近邻子空间法(NS)是一种基于信号重构的无参数分类器, 其分类思想是将测试样本表示为各类所有训练样本的线性组合, 从中选择最优解来进行分类. 2.1 表情数据库Cohn-Kanade数据库含有210个对象的大约2000个左右的具有充足的正面光照的灰度图像序列. 图像序列的分辨率都是640×490. 该数据库总共含有七种基本的表情, 如生气, 高兴、悲伤、惊奇、讨厌、害怕以及中性, 如图1所示. 我们从数据库中选用来自96个对象的320图像序列用于实验测试. 选择图像序列的标准是能够标记出除中性之外的六种表情. 然后对每个选择的图像序列中提取出一帧中性表情图像以及一帧含六种表情之一的图像. 最后我们提取出包括七种表情的470幅图像, 其中生气32个, 高兴100个, 悲伤55个, 惊奇75个, 害怕47个, 讨厌45个和中性116个.2.2 无腐蚀的人脸表情识别实验在该实验中, 直接使用32×32像素大小的图像样本用于表情识别, 图像中不存在任何腐蚀现象. 表1列出了SRC、NN, SVM和NS四种不同方法所取得的人脸表情识别性能. 由表1可知, 在无任何腐蚀图像的条件下, 稀疏表示分类方法SRC取得的人脸表情识别性能最好, 达到94.76%的识别率. 这表明了SRC用于人脸表情识别具有优越的分类性能.为了进一步给出七种表情中不同表情的具体识别性能, 表2给出了在Cohn-Kanade数据库上SRC方法采用32×32像素所取得的不同表情的识别结果. 从表2的实验结果可见, 在Cohn-Kanade数据库上七种表情中大部分表情的正确识别率达到了100%.2.3 有腐蚀的人脸表情识别实验为了检验SRC的鲁棒性人脸表情识别性能, 对32×32像素大小的测试图像随机添加像素腐蚀(Pixel Corruption). 随机添加像素腐蚀就是从测试图像中随机选择一定比例的像素, 采用范围之内的随机值进行替代, 其中表示第个测试图像的最大像素值. 实验中, 像素腐蚀比例从0%到90%, 依次递增10%. 图2展示了Cohn-Kanade数据库中一副原始图像从采样到腐蚀的过程, 其中图(a)为原始640×490像素的图像, 图(b)为采样之后的32×32像素的图像, 图(c)对32×32像素图像添加50%的腐蚀比例之后的图像.图3列出了NN、SVM、NS和SRC四种方法在Cohn-Kanade数据库上随机添加像素腐蚀比例从0%到90%取得的识别结果. 由图3实验结果可见, 随着图像腐蚀比例的增大, 图像越来越模糊, 人脸表情识别率也随之下降. 在图像腐蚀比例由0%增长到30%为止, SRC的正确识别率下降速度缓慢, 而其他三种方法的识别率下降非常快. 随之腐蚀比例的不断增大(30%至90%), 各种方法的识别率都一致下降, 但是SRC方法的识别率平均超过其它三种方法10%以上. 显然, 我们看到了SRC方法在处理人脸表情问题上有着良好的鲁棒性. 这主要是SRC方法提取了信号的稀疏结构, 并利用l1范数来作为来求解信号的稀疏表示系数. 由于采用正则化技术, SRC 的稀疏表示系数具有非常稳定的数值解.本文通过考虑测试图像是否存在像素腐蚀的现象, 对基于压缩感知理论的稀疏表示分类方法SRC的鲁棒性人脸表情识别性能进行了探讨. 在无任何像素腐蚀的人脸表情识别实验中, SRC取得的人脸表情识别性能比其他方法高出2%左右, 而在有像素腐蚀图像的人脸表情识别实验中, SRC展示出了良好的鲁棒性性能, 尤其在像素腐蚀比例30%至90%之间, SRC比其他方法的识别率平均高出10%以上. 这表明本文采用的基于压缩感知理论的稀疏表示分类方法SRC用于鲁棒性人脸表情识别时, 拥有良好的分类性能和鲁棒性.1 Tian Y, Kanade T, Cohn JF. Facial expression recognition. Handbook of Face Recognition, 2011: 487–519.2 刘晓旻,章毓晋.基于Gabor直方图特征和MVBoost的人脸表情识别.计算机研究与发展,2007,44(7):1089–1096.3 刘帅师,田彦涛,万川.基于Gabor多方向特征融合与分块直方图的人脸表情识别方法.自动化学报,2012,37(12): 1455–1463.4 易积政,毛峡,薛雨丽.基于特征点矢量与纹理形变能量参数融合的人脸表情识别.电子与信息学报,2013,35(10): 2403–2410.5 朱晓明,姚明海.基于局部二元模式的人脸表情识别.计算机系统应用,2011,20(6):151–154.6 Aleksic PS, Katsaggelos AK. Automatic facial expression recognitionusing facial animation parameters and multistream HMMs. IEEE Trans. on Information Forensics and Security, 2006, 1(1): 3–11.7 Zheng W, Zhou X, Zou C, et al. Facial expression recognition using kernel canonical correlation analysis (KCCA). IEEE Trans. on Neural Networks, 2006, 17(1): 233–238.8 Zhao G, Pietikainen M. Dynamic texture recognition using local binary patterns with an application to facial expressions. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2007, 29(6): 915–928.9 Zhao X, Zhang S. Facial expression recognition using local binary patterns and discriminant kernel locally linear embedding. EURASIP Journal on Advances in Signal Processing, 2012, (1): 20.10 Yurtkan K, Demirel H. Feature selection for improved 3D facial expression recognition. Pattern Recognition Letters, 2014, 38: 26–33.11 Candes EJ, Wakin MB. An introduction to compressive sampling. IEEE Signal Processing Magazine, 2008, 25(2): 21–30.12 Donoho DL. Compressed sensing. IEEE Trans. on Information Theory, 2006, 52(4): 1289–1306.13 Yang J, Wright J, Huang TS, et al. Image super-resolution via sparse representation. IEEE Trans. on Image Processing, 2010, 19(11): 2861–2873.14 Wright J, Yang AY, Ganesh A, et al. Robust face recognition via sparse representation. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2009, 31(2): 210–227.15 Mei X, Ling H. Robust Visual Tracking and Vehicle Classification via Sparse Representation. IEEE Trans. on Pattern Analysis and MachineIntelligence, 2011, 33(11): 2259–2272.16 Kanade T, Tian Y, Cohn J. Comprehensive database for facial expression analysis. International Conference on Face and Gesture Recognition. Grenoble, France. 2000. 46–53.17 Lee KC, Ho J, Kriegman DJ. Acquiring linear subspaces for face recognition under variable lighting. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2005, 27(5): 684–698.。
于慧敏,浙江大学,教授,博士生导师。主要研究方向为图像视频处理与
于慧敏,浙江大学,教授,博士生导师。
主要研究方向为图像/视频处理与分析。
2003年获科学技术三等奖一项,授权发明专利近20项,多篇论文发表在模式识别和计算机视觉领域顶尖学报和会议上。
近年来,在 (3D/2D)视频/图象处理与分析、视频监控、3D视频获取和医学图像处理等方面,主持了多项国家自然科学基金、973子课题、国家国防计划项目、国家863课题、浙江省重大/重点项目的研究和开发。
一、近年主持的科研项目(1)国家自然基金,61471321、目标协同分割与识别技术的研究、2015-2018。
(2) 973子课题,2012CB316406-1、面向公共安全的跨媒体呈现与验证和示范平、2012-2016。
(3)国家自然基金,60872069、基于3D 视频的运动分割与3D 运动估计、2009-2011。
(4) 863项目,2007AA01Z331、基于异构结构的3D实时获取技术与系统、2007-2009。
(5)浙江省科技计划项目,2013C310035 、多国纸币序列号和特殊污染字符识别技、2013-2015。
(6)浙江省科技计划重点项目, 2006C21035 、集成化多模医学影像信息计算和处理平台的研发、2006-2008。
(7)航天基金,***三维动目标的获取与重建、2008-2010。
(8)中国电信,3D视频监控系统、2010。
(9)中兴通讯,跨摄像机的目标匹配与跟踪技术研究、2014.05-2015.05。
(10)浙江大力科技,激光雷达导航与图像读表系统、2015-。
(11)横向,纸币序列号的实时识别技术、2011-2012。
(12)横向,清分机视频处理技术、2010-2012。
(参与)(13)横向,基于多摄像机的目标跟踪、事件检测与行为分析、2010。
(14)横向,红外视频雷达、2010-2012。
(15)横向,客运车辆行车安全视频分析系统、2010-2011。
二、近五年发表的论文期刊论文:1)Fei Chen, Huimin Yu#, and Roland Hu. Shape Sparse Representation for JointObject Classification and Segmentation [J]. IEEE Transactions on Image Processing 22(3): 992-1004 ,2013.2)Xie Y, Yu H#, Gong X, et al. Learning Visual-Spatial Saliency for Multiple-ShotPerson Re-Identification[J].Signal Processing Letters IEEE, 2015, 22:1854-1858.3)Yang, Bai, Huimin Yu#, and Roland Hu. Unsupervised regions basedsegmentation using object discovery, Journal of Visual Communication and Image Representation, 2015,31: 125-137.4)Fei Chen, Roland Hu, Huimin Yu#, Shiyan Wang: Reduced set density estimatorfor object segmentation based on shape probabilistic representation. J. Visual Communication and Image Representation,2012, 23(7): 1085-1094.5)Fei Chen, Huimin Yu#, Jincao Yao , Roland Hu ,Robust sparse kernel densityestimation by inducing randomness[J],Pattern Analysis and Applications: Volume 18, Issue 2 (2015), Page 367-375.6)赵璐,于慧敏#,基于先验形状信息和水平集方法的车辆检测,浙江大学学报(工学版),pp.124-129,2010.1。
典型室内场景显著性稀疏识别
典型室内场景显著性稀疏识别严晗;刘佶鑫;龚建荣【摘要】With the development and popularization of information technology and intelligent robots,scene recognition as an important research content has become an important research in the field of computer vision and pattern recognition problem.Solving the problem of the low classification accuracy for indoor scene will help the indoor scene classification in some areas of application:the image retrieval,video retrieval of the scene and the robot.Conventional scene recognition methods have poor performance in indoor situations.For this reason,a sparse representation indoor scene recognition method is presented,which based on significant detection.This method is using significant recognition detection to extract the scene in the image area which we are interested in,and combined with sparse representation to scene classification recognition.Experimental results show that this method can be applied to a typical family indoor scenarios (e.g.,bedroom,kitchen,closet,etc.)and have certain advantages in terms of recognition accuracy.%随着信息技术和智能机器人的发展与普及,场景识别作为重要的研究内容,已成为计算机视觉和模式识别领域的重要研究问题.解决室内场景分类精度低的问题,将有助于室内场景分类在场景图片检索、视频检索及机器人等领域中的应用.针对常规场景识别方法在室内环境中性能显著下降的问题,提出一种基于显著性检测的稀疏表示室内场景识别方法.该方法利用显著性区域检测算法提取出场景图像中人眼感兴趣的区域,并与稀疏表示结合进行场景识别.实验结果表明,将本方法应用在典型家庭室内场景(如卧室、厨房、衣帽间等),在识别正确率方面有一定的优势.【期刊名称】《南京师大学报(自然科学版)》【年(卷),期】2017(040)001【总页数】7页(P79-85)【关键词】场景识别;室内场景分类;显著性区域检测;稀疏表示【作者】严晗;刘佶鑫;龚建荣【作者单位】南京邮电大学通信与信息工程学院,江苏南京210003;南京邮电大学教育部工程研究中心,江苏南京210003;南京邮电大学教育部工程研究中心,江苏南京210003【正文语种】中文【中图分类】TP391由于多媒体技术和互联网技术的快速发展,每天都会产生大量的数字图像. 如何利用计算机自动将图像按照人类理解的方式分类到不同的类别,从而快速有效地获取、管理和分类数量巨大的图像成为一个重要问题,场景分类就由此产生. 针对场景分类难的问题,不同阶段的研究提出不同的方法和模型. 早期,场景图像分类一般都是基于整体模型[1-2]开展,利用颜色、纹理、形状等特征进行识别,该类方法在训练集之外的泛化能力较差. 近年来,诸如SIFT[3,19](Scale-invariant Feature Transform,尺度不变特征变换)、SURF[4](Speeded-Up Robust Features,加速鲁棒特征)、HOG[5](Histogram of Oriented Gradient,梯度方向直方图)、SDA[6](Subclass Discriminant Analysis,子类判别分析)等算子有着广泛应用. 比较流行的分类方法是利用各种算子提取环境特征,应用最广的是视觉词袋模型(Bag of visual Words,BoW[7]),这一思路在图像分类的应用中取得了令人鼓舞的结果,受到了研究者的极大关注. 但是,由于忽略了局部图像块的位置信息,该方法属于一种无序的特征表示,即缺少位置信息的全局特征表示. 为解决这个问题,Lazebnik等人提出以一种基于空间金字塔匹配(Spatial Pyramid Matching,SPM[8,18])的方法来改进传统的视觉词袋模型. 但是SPM策略存在着较大的量化误差,进而导致比较严重的信息损失. 为了解决这个问题,Yang J[9] 等人首次提出使用稀疏编码(Sparse Coding)策略来学习视觉词典,然后用稀疏编码方法对整幅图像的关键点进行编码,最后用基于最大池(Max Pooling)的方法表示图像特征,他们称这种方法为稀疏的空间金字塔匹配(Sparse coding Spatial Pyramid Matching,ScSPM). 此后,又有一系列工作对ScSPM中的不足做了改进,如Wang[10]等人在稀疏编码中加入了位置信息的约束,这使得编码效率和性能得到了改善;Boureau[11]等人在视觉学习中引入了类别信息,提出了监督的系数词典学习方法.室内场景分类是场景分类的一个研究领域,解决室内场景分类精度低的问题,将有助于室内场景分类在场景图片检索、视频检索及机器人等领域的应用. 但是,现在所有的算法只是对室外场景处理较好,对于室内场景的识别还存在很多的不足. 这是因为相较于户外场景,室内环境通常缺少显著的局部或全局视觉特征. 本文针对常规场景识别方法在室内环境中性能显著下降的问题,提出一种基于显著性检测的稀疏表示室内场景识别方法. 该方法利用显著性区域检测提取出场景图像中人眼感兴趣的区域,并与稀疏表示结合进行场景识别. 实验结果表明,本方法能得到较高的识别正确率.场景识别技术的典型框架是特征表示加分类器,其中的特征表示算子都是人为设计的,需要有特征提取的预处理过程.本文的算法框架如图1所示. 从框架流程图中可以看出,该室内场景识别算法结合了显著性检测和稀疏表示算法,场景图片首先通过显著性检测得到图像的显著性图,通过该显著性图得到不同的分割图像,这些分割图像作为最终稀疏表示算法的训练输入,最终训练出类别字典进行场景识别. 显著性检测不仅分割出图像中最主要的目标,减少了背景噪声的干扰,而且提高了稀疏表示算法的运算速度和可靠性.1.1 显著性检测算法当浏览一个场景的时候,人类视觉具有倾向于忽略不重要的区域而快速地搜索到感兴趣目标的能力. 这些区域通常被大家称为奇异点、视觉焦点或者是显著性区域[12]. 图像显著性区域检测能够帮助大家采用不同的策略处理不同的区域. 例如,通常采用精确的方法处理显著性区域,采用近似的方法处理非重要的区域. 通过这种方式,避免了对整幅图像应用复杂的算法,从而提高了图像处理速度.最早提出来的显著性算法是基于生物启发模型. 这个方法模拟人的视觉神经,通过计算“center-surround difference”来获取到显著性对象的位置信息. 但是由于该方法只计算局部特征的对比度,得到的显著度图只高亮了对象的边缘信息. 因此,后来的显著性区域检测算法主要集中尝试利用各种策略避免结果中只高亮边缘信息. 当前显著性检测的方法有很多,其中最为典型的方法有基于局部对比分析的算法、基于图论的算法、基于频谱特性的SR(Spectrum Residual)算法等[13],这些方法遵循的视觉显著性规律不同、使用的图像特征不同、对特征的处理方式也不同,所得到的显著图也都有自己的特点,但从所有方法中都可以发现设计显著性检测方法的基本思路.本文使用的显著性目标检测方法是由Huaizu、 Jiang等人在2013年提出的不同区域特征融合(Discriminative Regional Feature Integration,DRFI)的显著性检测方法[14].该方法有3个主要步骤组成,包括多尺度分割、区域显著性计算和多尺度显著性融合. 其原理如图2(详见文献[14])所示.1.1.1 多尺度分割多尺度分割中采用基于图的图像分割方法. 给一张原始图片I,将其进行M尺度的分割得到S={S1,S2,…,SM},其中每个SM都是原始图片I的分割,包括Km个区域. S1是最好的分割,它包含了最大数量的分割区域,SM是最粗糙的分割,它拥有最少的分割区域.1.1.2 显著性分数计算和大多数显著性检测方法不同的是,该显著性算法是先设计出一些对比特征向量,然后将这些有效的对比特征向量作为训练特征来训练一个随机森林回归器,该回归器负责计算不同图片的显著性分数.显著性特征向量分为3个主要部分:区域对比描述、区域属性描述和区域背景特征描述. 其中区域对比描述主要包括每张分割图中相邻区域之间颜色和纹理之间的差异,包括RGB均值、L*a*b均值、LM滤波器绝对响应、LM滤波器的最大响应、L*a*b直方图、灰度直方图、饱和度直方图和纹理直方图. 其中一般特征向量之间的差异以如下形式来计算:而直方图之间的差异计算方式为:区域属性描述又包括外表属性描述和集合属性描述,外表特征视图描述图像区域的颜色和纹理特征,它们可以作为区分显著性区域和背景的最基本的属性. 几何特征包括大小和位置关系,这些对描述显著性和背景之间的空间关系有一定影响.1.1.3 多尺度显著性融合多尺度显著性融合的目的就是将多尺度分割后计算得到的显著性图融合成一张显著性图,该方法使用了一个线性融合器ωmAm来进行融合操作,这个线性融合器通过使用最小化均方误差估来学习参数,即最小化目标函数:1.2 稀疏表示分类算法稀疏表示(Sparse Representation)理论是一种新兴的信号表示方法,此方法使用超完备字典对信号进行分解,对信号的误差与噪声比传统方法更稳健. 在图像压缩领域中,更稀疏的字典能够得到更高的压缩比;在图像重建领域中,更稀疏意味着更高质量的图像重建. 由此可见,稀疏性对于图像表示(Image Representation)是至关紧要的. 另外,基于稀疏分类框架[15-17]的目标跟踪算法和图像分类算法具有独特的抗噪声与遮挡能力.信号的稀疏表示是数据表示体系的重要组成之一. 设字典A由一组线性独立的基矢量[a1,a2,…,aM](原子)组成,这些基矢量能够张成整个矢量空间X=[x1,x2,…,xM]∈RM,即空间中任意矢量x都可以通过这组基的线性组合进行重构,如式(4)所示:式中,ci=〈ai,x〉是x在基矢量ai上的展开系数. 因为基是相互独立的,则这种展开的结果是唯一的. 如果ai⊥aj,则字典A为空间X的一组正交基. 则式(4)可改写为式(5):在过完备字典上求解稀疏表示问题可以用1范数最小化方法来解决:在式(6)中,A代表训练字典,y表示测试样本,x表示稀疏系数.在经典的稀疏表示算法中,稀疏表示拟解决的问题可以表示如式(7)所示:式中,y表示测试样本,X表示训练字典,α则表示测试样本y在训练字典X下的系数. 通过计算每一类的残差ei(y)=‖i‖2,并根据残差大小,从而可以判断测试样本所属的类别.稀疏表示方法是基于每一类字典Xi都是过完备的假设的. 但本文针对的居家室内场景识别分类,这是一个小样本问题,所得到的训练字典X也并不是完备的. 如果依旧用测试样本y的类字典Xi来重现样本,那么误差就会相当大,最终将会导致得到的残差ei(y)和‖α‖1不精确,从而严重影响分类结果.为解决上述问题,本文拟引用最小均方差准则下的协同表示分类来改进稀疏表示模型,从而大大提高算法的识别速度和分类效果.利用类与类之间的相关性,即某些第j类的样本可能对第i类的测试样本的表示有着重要意义. 因此,可利用字典中其他类的图像来扩充本类图像. 在文献[15]中,就是利用这种方法来解决小样本问题的.这样,实际上就是在1范数的限制下,利用字典X=[X1,X2,…,Xn]中所有的数据来协同表示测试样本y. 那么接近优化式(7)就可变为式(8):有关联表示为y在平面X的投影. 在SRC(sparse representation-based classifier)中,残差ei(y)=‖‖2用于分类,则可以导出:ei(y)=‖‖2=‖‖.在式(9)中,‖对于所有的类来说都是连续的,那么很显然真正起作用的其实是:如图3所示, i和j. 图3中显示了测试样本y在平面X的投影,可看出与平行,则: 式中,(χi, i)是χi和i之间的夹角,是和χi之间的夹角,根据式(11)可得出:从式(12)中,可以看出,在判断测试样本y是否属于第i类时,不仅需要考虑和χi之间的夹角大小,还需要兼顾到χi和i之间的夹角大小,正是这种双重标准使得分类变得更加高效和鲁棒.本文通过MATLAB平台进行仿真实验,验证本文算法的场景分类效果. 本文的实验分为四个部分:第一个部分为显著性检测算法使用不同阈值时室内场景分类结果的对比实验;第二个部分为使用单个特征与使用多个特征融合分类结果的对比实验;第三部分为前两部分实验最优结果的结合,即本文最终算法的实验结果;最后一部分是检测本文算法复杂度的实验,实验通过处理不同像素大小的图片所需要的cpu时间来检验算法的复杂度.前三部分的实验中,所采用的室内场景均分为卫生间、卧室、衣帽间、厨房、客厅这5个场景类别. 从每个场景类别中随机选择50张图片作为字典训练的样本,并随机选择10张图片作为测试样本.而最后一个实验中,分别选择卧室、厨房、衣帽间等5个室内场景图片中像素大小为256×256、640×480、1080×810的图片各60张,其中50张用作训练字典,10张用作训练样本.本文对实验中的室内场景图片有一定的要求,即室内场景图片要相对简洁,且图片中一定有每类场景的代表性主目标,例如卧室图片一定包含床、客厅图片一定包含沙发或椅子等.2.1 固定阈值的显著性检测算法实验本实验部分针对的是显著性检测算法中阈值的选取问题. 为了得到最佳参数,将显著性检测算法中的阈值分别设为0(即不设置阈值)、0.2、0.4、0.6及0.8,其中,阈值的大小即表示对图像的分割程度,阈值越大,则分割后剩余的图像特征越少,不设阈值则表示不对图像进行分割.将不同阈值下得到的显著图与稀疏表示算法结合后进行室内场景的分类. 通过比较室内场景分类的正确率,从而得到分类效果最好的显著性检测算法阈值. 图4为不同阈值情况下,典型室内场景分类的平均正确率.由图4可知,不同的阈值设定使得室内场景的分类正确率不同,且当显著性检测算法的阈值设为0.4的时候,室内场景的分类效果最好,正确率为52%. 因此,本文在最终算法中将把显著性检测算法的阈值设置为0.4.2.2 显著性区域特征与灰度图特征融合实验本实验部分将各个场景的灰度图特征、显著图(无阈值)特征以及灰度图与显著图的融合特征分别作为稀疏表示算法的输入进行场景的分类,并比较各个实验的分类正确率,如图5所示.由图5可知,用单个特征即灰度图特征与显著图特征作为稀疏表示算法的输入时,室内场景的分类正确率为46%,而用灰度图与显著图的融合特征作为算法的输入时,场景的识别正确率为56%,这表明使用融合特征在一定程度上可以提高室内场景的分类正确率,比使用单个特征的分类效果好. 因此,在本文的最终算法中,将采用特征融合的方式.2.3 多特征融合算法实验根据2.1章节与2.2章节的实验结果,本实验部分将显著性检测算法的阈值设为0.4,并将融合特征作为稀疏表示算法的输入,即本文的最终算法为:将阈值为0.4的显著图与灰度图及显著图(未设阈值)特征融合,并作为稀疏表示算法的输入进行场景分类. 通过实验可知,本文算法的分类正确率为62%,比2.1章节与2.2章节的最好正确率都有所提高. 可见本文算法能够提高室内场景的分类正确率,有一定的实用价值.图6为部分场景及其灰度图与显著图.2.4 算法复杂度分析本实验部分,将像素大小为256×256、640×480、1080×810的图片分别进行实验,实验中,显著性检测的阈值设置为0.4,并采用融合特征作为稀疏表示算法的输入. 通过比较不同像素大小的图片在实验时所需要的cpu时间来检验算法的复杂度,并通过柱状图直观展现. (其中,实验所用电脑配置为:Intel Core i5 3.20 GHz,4 GB内存. )如图7所示,可以看出,同一实验条件下,本文算法在处理不同像素大小的图片时所需的cpu时间基本不变,由此说明本文算法的时间复杂度很小.3 结语针对多数常规算法在室外环境表现良好、在室内环境性能下降的问题,本文提出一种基于显著性检测的稀疏表示室内场景识别方法,创新性地将稀疏表示方法应用在家庭室内场景的识别中. 该方法利用显著性检测算法找出场景图像中人眼感兴趣的区域,并通过设置阈值提取出主要目标区域,并将其与稀疏表示算法结合进行场景的分类. 在实验中,将显著性检测算法的阈值设为0.4,并使用了多特征融合的方法. 通过这种方式,得到了各个场景的代表性目标,避免了对整幅图像应用复杂的算法,从而提高了图像处理速度与室内场景分类的正确率.【相关文献】[1] VAILAVA A,JAIN A,ZHANG H J. On image classification:city vs. landscape[C]//IEEE Workshop on Content-Based Access of Image and Video Libraries.Piscataway,USA:IEEE,1998:3-8.[2] CHANG E,GOH K,SYCHAY G,et al. CBSA:content-based soft annotation for multimodal image retrieval using bayes point machines[J]. IEEE transactions on circuits and systems for video technology,2003,13(1):26-38.[3] 钱堃,马旭东,戴先中,等. 基于层次化SLAM的未知环境级联地图创建方法[J]. 机器人,2011,33(6):736-741.[4] 包加桐,宋爱国,郭晏,等. 基于SURF特征跟踪的动态手势识别算法[J]. 机器人,2011,33(4):482-489.[5] ZHANG H B,SU S Z,LI S Z,et al. Seeing actions through scene context[J]. IEEE visual communications and image processing,2013,8 575(VCIP):1-6.[6] BEKIOS-CALFA J,BUENAPOSADA J M,BAUMELA L. Robust gender recognition by exploiting facial attributes dependencies[J]. Pattern recognition letters,2014,36:228-234.[7] LI F F,PERONA P. A Bayesian hierarchical model for learning natural scene categories[C]//IEEE Computer Society Conference on Computer Vision and Pattern Recognition. USA:IEEE Computer Society,2005:524-531.[8] LAZEBNIK S,SCHMID C,PONCE J. Beyond bags of features:spatial pyramid matching for recognizing natural scene categories[J]. IEEE conference on computer vision and pattern recognition(CVPR),2006,2:2169-2178.[9] YANG B J,YU K,GONG Y,et al. Linear spatial pyramid matching using sparse coding for image classification[C]//IEEE Computer Scoiety Conference on Computer Vision and Pattern Recognition. USA:IEEE,2009:1794-1801.[10] WANG J,YANG J,YU K,et al. Locality-constrained linear coding for image classification[C]//IEEE Computer Society on Computer,Vision and Pattern Recognition. US:IEEE,2010:3360-3367.[11] BOUREAU Y L,BACH F,LECUN Y,et al. Learning mid-level features forrecognition[C]//IEEE Conference on Computer Vision and Pattern Recognition.USA:IEEE,2010:2559-2556.[12] GOPALAKRISHNAN V,HU Y,RAJAN D. Random walks on graphs to model saliency in images[C]//IEEE Conference on Computer Vision and Pattern Recognition.USA:IEEE,2009:1698-1705.[13] AVIDAN S,SHAMIR A. Seam carving for content aware image resizing[J]. ACM Transactions on Graphics,2007,26(3):10-16.[14] JIANG H,WANG J,YUAN Z,et al. Salient object detection:a discriminative regional feature integration approach[C]//IEEE Conference on Computer Vision and Pattern Recognition. USA:IEEE,2014:2083-2090.[15] WRIGHT J,YANG A Y,GANESH A,et al. Robust face recognition via sparse representation[J]. IEEE transactions on pattern analysis and machine intelligence,2009,31(2):210-227.[16] XUE M,HAIBIN L. Robust visual tracking and vehicle classification via Sparse representation[J]. IEEE transactions on software engineering,2011,33(11):2259-2272.[17] HAN A,JIAO J,ZHANG B,et al. Visual object tracking via sample-based adaptive sparse representation[J]. Pattern recognition,2011,44(9):2170-2183.[18] HUANG F X. Beyond bag of latent topics:spatial pyramid matching for scene category recognition[J]. 浙江大学学报(英文版),2015,16(10):817-828.[19] HAYAT M,KHAN S H,BENNAMOUN M,et al. A Spatial layout and scale invariant feature representation for indoor scene classification[J]. Computer science,2015.。
Classification Essay
Classification Essay|Printable version| Definition:In a classification essay, we organize things into categories and give examples of things that fit into each category. For example, if you choose to write about types of computers (PCs and servers), each of your developmental paragraphs will define the characteristics of a different computer type.Classification criteria:Before writing, it is necessary to decide on the classification criteria. We should think according to what properties we are going to classify things. The criteria must be discriminating and the emerging classes should be non-overlapping.In the sample essay about types of computers, the computers are classified according to their functions and capabilities, as:Sample essay analysistopic: 5 types of computers criteria: their functions and capabilities1. PC general use by a singlepersondesktop:permanentlaptop: portable2. Workstation used for 3D graphics,game developmentpowerfulmicroprocessor,additionalmemory andenhancedcapabilities3. Server used to provide servicesto other computershave powerfulprocessors, lotsof memory andlarge harddrives4. Main frame used in business enableshundreds ofpeople to worktogether5. Super computer used for jobs that takemassive amounts ofcalculatingvery powerfulOrganization:The introduction of a classification essay is quite straightforward. In the thesis statement, you mention that there are (number) types of (something) according to their (properties).In the developmental paragraphs, you need to define each type you mentioned in the thesis. You may also need to show the similarities and/or differences of these types. Giving examples would enable your readers to understand better.Language:The common transitions used while classifying are the first kind / type / group, the second kind / type / group, the third kind / type / group. Sample classification essay:Types of ComputersThere are a lot of terms used to describe computers. Most of these words imply the size, exp ected use or capability of the computer. While the term “computer”can apply to virtually any device that has a microprocessor in it, most people think of a computer as a device that receives input from the user through a mouse or keyboard, processes it in some fashion and displays the result on a screen. Computers can be divided into five according to the purpose they are used for and their capabilities.The most familiar type of microprocessor is the personal computer (PC). It designed for general use by a single person. While a Mac is also a PC, most people relate the term with systems that run the Windows operating system. PCs were first known as microcomputers because they were a complete computer but built on a smaller scale than the huge systems in use by most businesses. A PC can come in two types (three if we include the Personal Digital Assistants (PDAs) that differ from PCs not by the working policy but in appearance as well.): Desktop and laptop. The former is not designed for portability. The expectation with desktop systems is that you will set the computer up in a permanent location. Most desktops offer more power, storage and versatility for less cost than their portable brethren. On the other hand, the laptops - also called notebooks - are portable computers that integrate the display, keyboard, a pointing device or trackball, processor, memory and hard drive all in a battery-operated package slightly larger than an average hardcover book.Another purpose for using a microprocessor is as a workstation. The computers used for this purpose have a more powerful processor,additional memory and enhanced capabilities for performing a special group of task, such as 3D Graphics or game development.A computer can also be used as a server. For this, it needs to be optimized to provide services to other computers over a network. Servers usually have powerful processors, lots of memory and large hard drives.A fourth type, a main frame is the heart of a network of computers or terminals which allows hundreds of people to work at the same time on the same data. It is indispensable for the business world. Sometimes, computers can be used for specialized fields as well. T he supercomputer is the top of the heap in power and expense. It is used for jobs that take massive amounts of calculating, like weather forecasting, engineering design and testing, serious decryption, and economic forecasting.With the increasing demand in different specialties, new adjustments are being made to microprocessors and new types of computers that serve different purposes emerge. In this ongoing process, it would not possible to put a full stop here. What we suggest is that it is better to keep en eye on the development of science in this field and keep updating our knowledge in order not to be out-of-date like the computers of old times that were as big as a room.。
基于自适应权重的多重稀疏表示分类算法_段刚龙_魏龙_李妮
网络出版时间:2012-08-16 10:45网络出版地址:/kcms/detail/11.2127.TP.20120816.1045.019.htmlComputer Engineering and Applications计算机工程与应用基于自适应权重的多重稀疏表示分类算法段刚龙, 魏龙, 李妮DUAN Ganglong, WEI Long, LI Ni西安理工大学信息管理系, 陕西西安 710048Department of Information Management, Xi’an University of Technology, Xi’an 710048, ChinaAdaptive weighted multiple sparse representation classification approach Abstract:An adaptive weighted multiple sparse representation classification method is proposed in this paper. To address the weak discriminative power of the conventional SRC (sparse representation classifier) method which uses a single feature representation, we propose using multiple features to represent each sample and construct multiple feature sub-dictionaries for classification. To reflect the different importance and discriminative power of each feature, we present an adaptive weighted method to linearly combine different feature representations for classification. Experimental results demonstrate the effectiveness of our proposed method and better classification accuracy can be obtained than the conventional SRC method.Key words:adaptive weight; multiple sparse representation; SRC摘要:提出了一种基于多特征字典的稀疏表示算法。
模式识别论文(Pattern recognition)
模式识别论文(Pattern recognition)Face recognition based on sparse representationImage sparse representation of the image processing in the exergy is very suitable for image sparse representation of the image obtained by decomposition of gaugeThe calculations are enormous. Using MP implementation method based on image sparse decomposition algorithm using genetic algorithm for fast exergy processThe best atom is decomposed at each step.The problem of face recognition is a classical pattern recognition problem. In recent years by the Exergy Theory of compressed sensing based on dilute inspired exergySparse representation of face recognition technology has been extensively studied. Face recognition based on sparse representation is the construction of words using training picturesThe sparse linear combination coefficients and exergy exergy code by solving an underdetermined equation to obtain the test images according to these coefficientsThe image recognition classification.Keywords image processing in the sparse representation of the MP within the genetic algorithm of sparse decompositionFace, recognition, via, sparse, representationAbstract:, sparse, representation, of, images, is, very, suitable,, for, image, processing,But, the, computational, burden, in, sparse, decomposition, process, image, is, huge,, A, newFast, algorithm, was, presented, based, on, Matching, Pursuit (MP), image, sparseDecomposition. At, first, Genetic, Algorithms (GA), was, applied, to, effectively, searchIn, the, dictionary, of, atoms, for, the, best, atom, at, each,, step, of, MPFace, recognition, problem, is, a, classic, problem, of, pattern,, recognition., In, recentYears, inspired, by, the, theory, of, perception, is, compressed, sparseRepresentation-based, face, recognition, technology, has, been, widely, studied., FaceRecognition, based, on, sparse, representation, is, to, take, advantage,, of, the, trainingImages, constructed, dictionary, owed, by, solving, a, the, most,, sparse, linear, combinationCoefficients, given, equation, to, obtain, the, test, images, then, these, coefficients, toIdentify image classification.Key words: image processing; sparse representation; sparse decomposition;Matching Pursuit; Genetic Algorithms0 Introduction the current face recognition technology of rapid development especially the exergy basedStatic face detection and recognition, and face feature extractionMulti face recognition based on multi pose has been achievedA great deal of research. But the exergy exergy in more complex environmentsSuch as facial expression recognition, illumination compensation and Guang ZhaomoThe establishment of the model, the treatment of age changes, and a variety of testing dataThere is a lack of effective methods for fusion.Face recognition includes three steps in face detectionMeasurement, face feature extraction, face recognition and verification. There arePeople on thisExtension of the exergy based on the above three stepsOn Exergy increased early standardization, and correction and later pointsClass and management these two steps.The research of face recognition started in the late 1960sL2]. Has experienced 40 years of development. Roughly divided into threeThree stages:The first stage is the initial stage from 60s to the end of exergyLate 80s. The main technique adopted at that time was baseTo set the structure characteristics of the face recognition method of exergy isAs a general pattern recognition problem is studied. generationThe figures include Bledsoe (Bledsoe) and Gordon Stein(Goldstein), Harmon (Harmon), and Kim Wu Hsiung(KanadeTakeo) et al. At that time almost all were identifiedThe process relies on manual operation and results in no exergy into very important practical applications in not many basically noHave practical application.The second stage is in the exploration stage from 70s to eightThe ten age. During this period, as well as engineers in the smokeLead neuroscientists and psychologists to the fieldResearch. The former is mainly through the perception mechanism of the human brainTo explore the possibility in automatic face recognition while the orderSome theoretical obtained has some defects and partial nature but inEngineering techniques for design and implementation of algorithms and systemsThe personnel have the important theory instructionsignificance.The third stage is the stage of rapid development in the last century from the nineFrom the ten to the present. Computer vision and pattern recognition technologyIn the rapid development of computer image processing technology and drivesThe rapid development of face recognition. Governments are also heavily financedIn the study of face recognition and achieved fruitful results.Among them, Eigenfaee and Fisherface is this momentThe most representative, the most significant achievements of the twoThree kinds of face recognition algorithms have become the base of face recognitionAlgorithms and industrial standards.1 sparse representation of the mathematical form of sparse representation of the face recognition problem is represented mathematicallyF = A X Y is in the m where Y is the dimension of natural channelNo, A is also known from a predefined dictionary based X is a natural increase.The n-dimensional sparse representation of signals under predefined bases. KnownBased on the original signal by solving its in the predefined baseIn the sparse representation is a sparse encoding problem in the following twoSolution method]3-1 [fSparse encoding f sparse regularization constraints K||X|| S.T. ||AX-Y||argmin0?The 22 rate in XThe error constrained sparse encoding exergy in FRate of 220 ||AX-Y|| S.T. ||X||argmin?XType F XIs the original signal Y, under the predefined baseThe sparse representation coefficient of exergy is share error tolerance share K is sparseShare threshold 0||The || said in that the number of columns of 0l norm vector 0Number of elements.Sparse coding and compressed sensing reconstruction of signals haveThat rate and the minimum eight norm can be very goodRestructure。
稀疏贝叶斯控制稀疏度的参数
稀疏贝叶斯控制稀疏度的参数介绍稀疏贝叶斯是一种经典的机器学习算法,用于处理高维数据集。
在稀疏贝叶斯中,控制稀疏度的参数起着重要的作用。
本文将探讨稀疏贝叶斯算法及其参数对稀疏度的影响。
稀疏贝叶斯简介稀疏贝叶斯是基于贝叶斯理论的一种分类算法。
它假设每个特征都是独立的,并且每个特征的概率分布都是高斯分布。
稀疏贝叶斯通过引入稀疏先验分布来实现特征的选择,从而达到降低维度和提高模型泛化能力的目的。
稀疏度的定义稀疏度是指模型中非零特征的比例。
在稀疏贝叶斯中,稀疏度越高,表示模型选择的特征越少,模型的泛化能力越强。
稀疏度参数的选择稀疏贝叶斯中有两个重要的参数控制稀疏度,分别是超参数alpha和beta。
下面将详细介绍这两个参数的作用和选择方法。
超参数alpha超参数alpha用于控制特征的稀疏度。
较大的alpha值会使得模型选择更少的特征,从而增加稀疏度。
较小的alpha值会使得模型选择更多的特征,从而降低稀疏度。
选择合适的alpha值是很重要的。
如果alpha值过大,模型可能会选择过少的特征,导致欠拟合。
如果alpha值过小,模型可能会选择过多的特征,导致过拟合。
一种常用的选择方法是使用交叉验证,在一定范围内选择alpha值,通过评估指标(如准确率或F1值)选择最优的alpha值。
超参数beta超参数beta用于控制特征的共享性。
较大的beta值会使得模型选择更多共享特征,从而增加稀疏度。
较小的beta值会使得模型选择更少共享特征,从而降低稀疏度。
选择合适的beta值也是很重要的。
如果beta值过大,模型可能会选择过多共享特征,导致过拟合。
如果beta值过小,模型可能会选择过少共享特征,导致欠拟合。
同样,可以使用交叉验证来选择最优的beta值。
稀疏贝叶斯控制稀疏度的参数实验为了验证上述参数对稀疏度的影响,我们进行了一系列实验。
下面是实验的详细过程和结果。
数据集我们使用了一个经典的文本分类数据集,包含了多个类别的文本样本。
稀疏总结
稀疏表示在目标检测方面的学习总结1,稀疏表示的兴起大量研究表明视觉皮层复杂刺激的表达采用的是稀疏编码原则,以稀疏编码为基础的稀疏表示方法能较好刻画人类视觉系统对图像的认知特性,已引起人们极大的兴趣和关注,在机器学习和图像处理领域得到了广泛应用,是当前国内外的研究热点之一.[1]Vinje W E ,Gallant J L .Sparse coding and decorrelation in pri- mary visual cortex during natural vision [J].Science ,2000,287(5456):1273-1276.[2]Nirenberg S ,Carcieri S ,Jacobs A ,et al .Retinal ganglion cells act largely as independent encoders [J ].Nature ,2001,411(6838):698-701.[3]Serre T ,Wolf L ,Bileschi S ,et al .Robust object recognition with cortex-like mechanisms[J].IEEE Transactions on PatternAnalysis and Machine Intelligence ,2007,29(3):411-426.[4]赵松年,姚力,金真,等.视像整体特征在人类初级视皮层上的稀疏表象:脑功能成像的证据[J].科学通报,2008,53(11):1296-1304.图像稀疏表示研究主要沿着两条线展开:单一基方法和多基方法.前者主要是多尺度几何分析理论,认为图像具有非平稳性和非高斯性,用线性算法很难处理,应建立适合处理边缘及纹理各层面几何结构的图像模型,以脊波(Ridgelet)、曲波(Curvelet)等变换为代表的多尺度几何分析方法成为图像稀疏表示的有效途径;后者以Mallat 和Zhang 提出的过完备字典分解理论为基础,根据信号本身的特点自适应选取能够稀疏表示信号的冗余基。
基于BERT模型的中文短文本分类算法
第47卷第1期Vol.47No.1计算机工程Computer Engineering2021年1月January 2021基于BERT 模型的中文短文本分类算法段丹丹1,唐加山1,温勇1,袁克海1,2(1.南京邮电大学理学院,南京210023;2.圣母大学心理学系,美国南本德46556)摘要:针对现有中文短文本分类算法通常存在特征稀疏、用词不规范和数据海量等问题,提出一种基于Transformer 的双向编码器表示(BERT)的中文短文本分类算法,使用BERT 预训练语言模型对短文本进行句子层面的特征向量表示,并将获得的特征向量输入Softmax 回归模型进行训练与分类。
实验结果表明,随着搜狐新闻文本数据量的增加,该算法在测试集上的整体F 1值最高达到93%,相比基于TextCNN 模型的短文本分类算法提升6个百分点,说明其能有效表示句子层面的语义信息,具有更好的中文短文本分类效果。
关键词:中文短文本分类;基于Transformer 的双向编码器表示;Softmax 回归模型;TextCNN 模型;word 2vec 模型开放科学(资源服务)标志码(OSID ):中文引用格式:段丹丹,唐加山,温勇,等.基于BERT 模型的中文短文本分类算法[J ].计算机工程,2021,47(1):79-86.英文引用格式:DUAN Dandan ,TANG Jiashan ,WEN Yong ,et al.Chinese short text classification algorithm based on BERT model [J ].Computer Engineering ,2021,47(1):79-86.Chinese Short Text Classification Algorithm Based on BERT ModelDUAN Dandan 1,TANG Jiashan 1,WEN Yong 1,YUAN Kehai 1,2(1.College of Science ,Nanjing University of Posts and Telecommunications ,Nanjing 210023,China ;2.Department of Psychology ,University of Notre Dame ,South Bend 46556,USA )【Abstract 】The existing Chinese short text classification algorithms are faced with sparse features ,informal words andmassive data.To address the problems ,this paper proposes a Chinese short text classification algorithm based on the Bidirectional Encoder Representation from Transformer (BERT )model.The algorithm uses BERT pre -training language model to perform eigenvector representation of short text on the sentence level ,and then the obtained eigenvector is input into the Softmax regression model for training and classification.Experimental results show that with the growth of data from Sohu news ,the overall F 1value of the proposed algorithm on the test dataset is up to 93%,which is 6percentage points higher than that of the TextCNN -based short text classification algorithm.The result demonstrates that the proposed algorithm performs better in semantic information representation at the sentence level ,and in the classification of Chinese short texts.【Key words 】Chinese short text classification ;Bidirectional Encoder Representation from Transformer (BERT );Softmax regression model ;TextCNN model ;word 2vec model DOI :10.19678/j.issn.1000-3428.00562220概述根据中国互联网络信息中心于2019年2月28日发布的第43次《中国互联网络发展状况统计报告》[1],截至2018年12月我国网民规模达8.29亿,互联网普及率达到59.6%,其中网民通过手机接入互联网的比例高达98.6%,即时通信、搜索引擎和网络新闻是手机网民使用率最高的应用,这3类手机应用包含聊天记录、搜索日志、新闻标题、手机短信等大量短文本[2],携带了丰富的数据信息,其已成为人类社会的重要信息资源,如何高效管理这些海量的短文本并从中快速获取有效信息受到越来越多学者的关注,并且对于短文本分类技术的需求日益突显。
中药白英生药学鉴别研究
高师理科学刊Journal of Science of Teachers' College and University 第41卷第4期2021年 4月Vol. 41 No.4Apr. 2021文章编号:1007-9831 ( 2021 ) 04-0072-05中药白英生药学鉴别研究刘新波,吴静(连云港师范高等专科学校海洋港口学院,江苏连云港222006)摘要:采用石蜡切片技术、水合氯醛透化法对中药白英的营养器官解剖特征及组织粉末特征进行 研究.白英解剖结构、粉末特征显著,可见螺纹导管、网纹导管、含晶纤维、砂晶、环式茎表皮 气孔、不定式叶表皮气孔、腺毛、非腺毛.白英营养器官解剖结构、粉末特征可为白英生药鉴定 提供解剖学依据.关键词:白英;解剖结构;粉末特征;显微鉴别中图分类号:Q944 文献标识码:A doi : 10.3969/j.issn.1007-9831.2021.04.015Study on pharmacognostic identification of Solanum lyratumLIU Xinbo, WU Jing(School of Marine Port, Lianyungang Teacher 's College, Lianyungang 222006, China )Abstract : The anatomical characteristics of vegetative organs and the characteristics of tissue powder of Solanum lyratum w ere studied by paraffin section and chloral hydrate method. The anatomical structure and powder characteristics of Solanum lyratum were obvious. There were thread duct, reticulate duct, crystal fiber , sand crystal , circular stem epidermis stomata , adventitious leaf epidermis stomata , glandular hairs and non-glandular hairs. The anatomical structure and powder characteristics of the vegetative organs of Solanum Lyratum can be used as the basis for pharmacognosy identification.Key words : Solanum lyratum ; anatomical structure ; powder characteristics ; microscopic identification白英(Solanum lyratum Thunb .)为茄科多年生草质藤本植物,又名白毛藤、毛风藤、山甜菜、白石英.全 草入药,具有解毒、清热利湿、袪风化痰的功效,其根对于风湿性关节炎的治疗效果明显[1],果实能治风 火牙痛.白英全草中的主要有效成分为生物碱类、苷类、有机酸等化合物[2-3].药理研究表明,白英具有抑 制癌细胞的功效,临床上多用于治疗肝癌、胃癌、肺癌、乳腺癌、前列腺癌等[4-7].其还有镇痛抗炎冏、抗 菌、保肝、抗过敏等作用[9],可治疗胆结石、胆囊炎、风湿性关节炎等疾病[10].目前,国内对白英的研究主 要涉及药理学、化学成分、临床应用等方面,关于营养器官解剖结构、粉末特征的研究较少,且不全面.本 文对白英的营养器官解剖结构、粉末特征进行研究,以期为白英生药鉴定及质量标准制定提供依据.1材料与方法实验材料采自连云港市东磊村,经邵世光教授鉴定为茄科植物白英.石蜡制片:根、茎、叶直接来自所采野生白英植株,茎为当年生枝.根、茎切成约5 mm 小段,将叶 片沿主脉切成4 mmx6 mm 小块,置于FAA 固定液中固定24 h 以上.常规石蜡切片法切片,切片厚度10收稿日期:2021-01-04作者简介:刘新波( 1965-),女,黑龙江绥滨人,教授,从事植物学研究.E-mail : ***********************第4期刘新波,等:中药白英生药学鉴别研究73番红-固绿对染,中性树胶封片.临时装片:取带有叶脉的白英叶片,将其修整成1 emxl cm 小块,放入家用漂水中浸泡,至叶片颜色 褪去,反复冲洗后,用50%甘油水溶液制成临时装片,使用光学显微镜成像系统观察拍照.粉末装片:将干燥的白英根、茎、叶分别粉碎,过60目筛,采用常规水合氯醛透化法制片,光学显微 镜观察其粉末特征并拍照.2实验结果图 1 白英根横切结构2.1白英组织切片特征2.1.1白英根的解剖结构 白英根的横切面呈圆形,次生结 构由外至内为周皮、皮层、次生韧皮部、维管形成层、次生 木质部.周皮较发达,木栓形成层由1层细胞组成,向外分 裂形成木栓层,向内产生栓内层.木栓层细胞壁木栓化,栓 内层由1层细胞构成.次生韧皮部排列在木质部外方,次生 韧皮部与周皮所占比相差不多.根次生结构中的次生木质部 约占根横切面的1/2,导管孔径大小不均等(见图1).2.1.2白英茎的解剖结构 白英茎细长,不能直立,表面布满白色绒毛,横切面呈类圆形(见图2a ).表皮细胞位于茎 注:比例尺为100 p m (下同).Pd 为周皮;Co 为皮层;Ca 为 的最外侧,由1层细胞构成,排列紧密无缝隙,细胞体积普 形成层;为木质部;s p h 为次生韧皮;Pld 为栓内层.遍较小,具有保护作用,是茎的保护组织.表皮上具有表皮毛.皮层位于表皮和中柱之间,皮层细胞排列 疏松,可见明显的间隙,形状较规则,大多为薄壁细胞.皮层可分为内皮层、中皮层、外皮层3部分.外 皮层由2层紧密排列的细胞构成,内皮层细胞与外皮层相似,仅有1层,中皮层细胞层数较多,且细胞较 大,空隙明显(见图2b ).中柱在皮层的内部,也称为维管柱,由3部分构成,分别是维管束、髓、髓射 线.中柱近似圆形,维管组织由5~9个维管束组成,以环形方式排列,随着木质化程度增高,维管束则全 部相连,初生木质部在初生韧皮部内侧.髓位于茎的中央部分,在生长过程中毁坏形成空腔,髓由薄壁细 胞构成,细胞排列较疏松,形态较大,是维管柱中的基本组织.白英茎中的无固定形状晶体结构即砂晶[11], 主要存在于髓薄壁细胞和皮层细胞中(见图2e ).其形成的原因和功能有待进一步的研究.a 白英茎的横切面b 白英茎局部横切结构 e 白英茎中的晶体图2白英茎的横切结构注:Co 为皮层;PX 为初生木质部;PPh 为初生韧皮部;Pi 为髓;Ep 为表皮;EH 为表皮毛;Cr 为晶体.2.1.3白英叶的解剖结构 白英叶呈琴形或戟形,端钝,顶端稍尖•叶长3〜5 cm ,宽2〜5 cm ,两面均密生 白色绒毛•叶片厚度178~184.5 pm.叶的上、下表皮细胞形状不规则,但排列十分紧密,其垂周壁形成波 纹状弯曲,上表皮细胞波状弯曲较下表皮细胞浅(见图3a~b ).叶的上、下表皮均可见气孔器,气孔器随 机分布,上表皮气孔器很少,下表皮气孔明显多于上表皮,上、下表皮的气孔指数分别为0.048, 0.221, 由气孔细胞与表皮细胞的相对位置比较看,上表皮气孔的保卫细胞多与表皮细胞平行,下表皮气孔的保卫 细胞与其它表皮细胞不在一个平面,多向外隆起(见图3c ),气孔类型为不定式.上、下表皮外均有数量 较多的腺毛和非腺毛,腺毛有长腺毛和短腺毛2种.长腺毛长106〜1 686 pm,头部常由单细胞组成,柄部 由多细胞构成;短腺毛长88~112 pm ,头部由2~8个细胞组成,腺毛柄多为1个细胞.非腺毛多由1〜6个74高师理科学刊第41卷单列细胞组成,长度差别较大(见图3d~e).叶肉位于上、下表皮之间,由栅栏组织和海绵组织组成,栅栏组织由1列长圆柱形细胞组成,排列整齐且紧密.海绵组织由3~4层类圆形细胞组成•叶主脉近、远轴面均向外凸起•叶脉由机械组织和维管组织组成.主脉维管束1束,新月形,韧皮部分布于木质部的内外两侧,为双韧维管束(见图3f).叶柄由外至内为表皮、机械组织、基本薄壁组织、维管柱.表皮细胞近圆形,体积较小且排列紧密.机械组织与表皮紧密相连,由厚壁细胞组成•再往内是大量薄壁细胞•维管柱位于叶柄的中央,与主叶脉维管束相同•在叶柄近轴面端部有两条突起的纵棱脊,在棱脊下各有一束较小的维管束(见图3g).白英叶中具晶体,与茎中砂晶颜色相近,主要分散存在于海绵组织中(见图3h).a白英叶上表皮正面观b白英叶下表皮正面观c白英叶横切面d白英叶非腺毛、短腺毛横切面e白英叶长腺毛、非腺毛横切面f白英叶过主脉横切面g白英叶柄横切面h白英叶砂晶横切面图3白英叶片的横切结构注:C。
局部遮挡条件下的人脸识别算法研究
哈尔滨理工大学工学硕士学位论文局部遮挡条件下的人脸识别算法研究摘要随着人工智能的飞速发展,身份认证技术是现在重要的研究课题,而其中人脸识别技术作为一种新兴技术,深受许多领域关注。
随着人脸识别技术的发展,不断地有新型算法涌入,使得在受控条件下的人脸识别精度越来越高。
然而,在实际生活中,人脸识别技术常常面临着一些难题。
比如遮挡、光照、人脸表情与姿势等影响,造成人脸识别的效果不尽如人意。
其中局部遮挡会引起人脸识别的准确率急剧下降,因此本文针对局部遮挡的人脸识别展开讨论,通过所提算法提高了局部遮挡人脸识别的准确率:首先介绍人脸图像预处理技术中的图像增强技术与图像去噪技术,又介绍了卷积神经网络的理论,提出了针对局部遮挡人脸的LLE-CNN人脸检测方法,该方法可以检测出局部遮挡的人脸并能够标示出遮挡部分。
其次根据对有遮挡人脸识别算法的分类,介绍了传统子空间方法。
传统子空间方法中稀疏表示分类方法与协同表示分类方法是针对局部遮挡人脸识别的两种常用算法,后又因为稀疏表示分类方法在实际中不足以描述编码错误,提出了一种稀疏表示分类方法的改进算法,鲁棒稀疏表示,并在AR人脸数据库中对比了三种算法的性能。
最后,为了提高局部遮挡人脸识别的鲁棒性。
本文采用了一种基于深度特征字典表示分类的算法,该算法利用卷积神经网络作为特征提取器,然后使用字典对提取的深度特征进行线性编码,最后利用稀疏表示编码的改进算法鲁棒稀疏编码进行残差分类。
通过实验证实该算法可以同时在有遮挡与无遮挡条件下的人脸识别都具有鲁棒性,在计算上是有效的,并且对于人脸图像中的较大连续遮挡也具有鲁棒性。
关键词人脸识别;遮挡人脸;子空间学习方法;字典学习;卷积神经网络-I-哈尔滨理工大学工学硕士学位论文Research on Face Recognition Algorithm underPartial OcclusionAbstractWith the rapid development of artificial intelligence,identity authentication technology is an important research topic now.Among them,face recognition technology,as an emerging technology,has attracted much attention in many fields. With the development of face recognition technology,new algorithms are constantly flowing in,which makes the accuracy of face recognition under controlled conditions getting higher and higher.However,in real life,face recognition technology often faces some difficulties.For example,the effects of occlusion,lighting,facial expressions,and postures cause the effect of face recognition to be unsatisfactory. Partial occlusion will cause the accuracy of face recognition to drop sharply. Therefore,this article starts a discussion on partial occlusion face recognition.The proposed algorithm improves the accuracy of local occlusion face recognition: Firstly,the image enhancement technology and image denoising technology in face image preprocessing technology are introduced,and the theory of convolutional neural network is also introduced.An LLE-CNN face detection method for partially occluding faces is proposed.This method can detect Faces that are partially occluded and can be marked with occlusion.Secondly,based on the classification of occluded face recognition algorithms, the traditional subspace method is introduced.The sparse representation classification method and collaborative representation classification method in the traditional subspace method are two commonly used algorithms for local occlusion face ter,because the sparse representation classification method is not sufficient to describe coding errors in practice,a sparse representation classification is proposed The improved algorithm,robust sparse representation,compares the performance of the three algorithms in the AR face database.Finally,in order to improve the robustness of local occlusion face recognition. This paper proposes an algorithm based on deep feature dictionary representation classification.This algorithm uses a convolutional neural network as a feature-II-哈尔滨理工大学工学硕士学位论文extractor,then uses a dictionary to linearly encode the extracted deep features,and finally uses an improved algorithm for sparse representation coding.Robust sparse coding Residual classification.Experiments show that the algorithm can be robust to face recognition under both occluded and unoccluded conditions,is computationally effective,and robust to large continuous occlusions in face images.Keywords Face recognition,occlusion face,Subspace learning method,dictionary learning,convolutional neural network-III-哈尔滨理工大学工学硕士学位论文目录摘要 (I)Abstract (II)第1章绪论 (1)1.1研究背景及意义 (1)1.1.1选题背景 (1)1.1.2解决人脸遮挡的现实意义 (2)1.2国内外研究现状 (2)1.2.1国外研究现状 (3)1.2.2国内研究现状 (5)1.3相关难点及问题的提出 (6)1.4本文主要研究内容 (6)第2章局部遮挡人脸识别综述 (8)2.1图像增强技术 (8)2.1.1图像灰度化处理 (8)2.1.2直方图均衡化 (9)2.2图像去噪技术 (10)2.2.1均值滤波 (11)2.2.2中值滤波 (11)2.2.3高斯滤波 (12)2.3卷积神经网络基础理论 (13)2.3.1卷积神经网络结构 (13)2.3.2VGG-Net基本结构 (19)2.3.3LLE-CNN人脸检测算法 (20)2.4本章小结 (22)第3章基于传统子空间方法的人脸识别方法研究 (23)3.1稀疏表示理论 (23)3.1.1稀疏表示 (24)3.1.2针对遮挡情况的稀疏表示分类算法 (25)3.1.3鲁棒稀疏编码模型 (26)哈尔滨理工大学工学硕士学位论文3.2基于协同表示的分类方法 (30)3.3实验结果与分析 (31)3.3.1AR人脸数据库简介 (31)3.3.2算法对比实验 (31)3.4本章小结 (33)第4章基于深度特征字典表示的遮挡人脸识别研究 (34)4.1深度特征字典表示法原理 (34)4.1.1利用卷积神经网络提取特征 (35)4.1.2字典表示 (36)4.1.3利用深度特征字典表示法处理遮挡 (37)4.1.4辅助字典的生成 (38)4.2实验结果与分析 (39)4.2.1不同遮挡物下的实验分析 (40)4.2.2单样本人脸识别 (41)4.2.3大型人脸数据库对算法的影响 (42)4.2.4深度特征字典表示法的计算复杂度 (43)4.2.5深度特征字典表示法算法性能评估 (43)4.3本章小结 (45)结论 (46)参考文献 (47)攻读硕士学位期间所发表的学术成果 (51)致谢 (52)哈尔滨理工大学工学硕士学位论文第1章绪论随着社会和经济的发展,人们进入信息化时代,互联网成为一种潜力无限的信息交互中心。
一种基于空域滤波的空间临近相干源角度估计方法
一种基于空域滤波的空间临近相干源角度估计方法郑轶松;陈伯孝;杨明磊【摘要】相干源常见于存在多径的场景,如何解相干历来是阵列信号处理领域亟待解决的难题之一,特别针对空间临近相干源,其角度估计精度尚有待提高。
针对空间临近相干源该文提出一种基于空域滤波的角度估计方法。
首先利用空域滤波技术将多个相干源分离,再对滤波分离后的各个信号分别进行角度估计,并通过对滤波器系数和相干源角度的迭代优化提高测角精度。
针对非均匀线阵,该方法采用虚拟阵列技术扩展其适用范围。
计算机仿真结果表明该方法的测角精度较现有方法更高,信噪比较高时其测角的均方根误差可达克拉美罗界,验证了该方法的有效性和在空间临近相干源场景的优越性。
%Coherent sources commonly exist in scenarios with multipath effect. How to decorrelate coherent sources is traditionally a problem urgently to be solved in the array signal processing domain. Especially for spatially adjacent coherent sources, the performance of the estimation of Direction Of Arrival (DOA) remains to be improved. A DOA estimation method based on spatial filtering is proposed for spatially adjacent coherent sources. Multiple coherent sources are separated by spatial filtering and the DOAs are estimated respectively afterwards. The performance of the DOA estimation is enhanced by refining the filter parameters and the DOAs of the coherent sources iteratively. To extend its application to non-uniform linear array, the virtual array technique is adopted. The computer simulation results indicate that the proposed algorithm has better DOA estimation performance than the existing methods. In the scenario of sufficiently high Signal to Noise Ratio(SNR), the Root Mean Square Error (RMSE) could achieve Cramer-Rao Bound (CRB). The effectiveness and the superiority of the proposed method for spatially adjacent coherent sources are validated by the simulation results.【期刊名称】《电子与信息学报》【年(卷),期】2016(038)012【总页数】7页(P3100-3106)【关键词】雷达信号处理;来波方向估计;空间临近相干源;空域滤波;解相干【作者】郑轶松;陈伯孝;杨明磊【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西安 710071; 西安电子科技大学信息感知技术协同创新中心西安 710071;西安电子科技大学雷达信号处理国家重点实验室西安 710071; 西安电子科技大学信息感知技术协同创新中心西安 710071;西安电子科技大学雷达信号处理国家重点实验室西安 710071; 西安电子科技大学信息感知技术协同创新中心西安 710071【正文语种】中文【中图分类】TN957.51空间临近相干源是指存在于一个波束宽度内的多个相干源目标,常见于存在多径干扰的关键场景,如雷达低仰角目标探测与跟踪[1,2]。
有关于分类的英语作文初一
When it comes to writing an essay about classification in English,especially for a seventhgrade student,its essential to approach the topic in a structured and clear manner. Here are some key points to consider when writing such an essay:1.Introduction:Begin by introducing the concept of classification.Explain that classification is a way of organizing information into groups based on shared characteristics.2.Importance of Classification:Discuss why classification is important.It helps in making sense of the world around us,whether its in biology,where animals and plants are classified into different species,or in everyday life,where we classify objects to keep our environment organized.3.Types of Classification:Describe the different types of classification.For example: Hierarchical Classification:This involves grouping items into a series of categories that are subdivided into smaller categories.Cross Classification:This is when items are classified based on multiple criteria. Binary Classification:This is a simple form of classification where items are divided into two categories.4.Examples in Daily Life:Provide examples of classification that students can relate to, such as:Organizing books in a library by genre and author.Sorting clothes by color or type.Classifying food into categories like fruits,vegetables,proteins,etc.5.Classification in Science:Explain how classification is used in various scientific fields. For instance:In Biology,the Linnaean system classifies organisms into kingdoms,phyla,classes, orders,families,genera,and species.In Geology,rocks are classified into three main types:igneous,sedimentary,and metamorphic.6.Classification in Technology:Discuss how technology uses classification,such as: Categorizing software applications based on their function.Organizing digital files and folders on a computer.7.Challenges in Classification:Mention some of the challenges that can arise when classifying,such as ambiguity in the criteria used for classification or the difficulty in categorizing items that dont fit neatly into one category.8.Conclusion:Sum up the essay by reiterating the importance of classification in organizing and understanding the world.Encourage students to think critically about the categories they use in their own lives and to consider how classification can be used to solve problems or make decisions.9.Personal Reflection:Optionally,you can include a personal reflection on how the student uses classification in their life or how they have learned to classify things in a new way.Remember to use simple and clear language appropriate for a seventhgrade level,and ensure that the essay is wellstructured with a logical flow of ideas.。
- 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)。
1 2
3
Numerical analysis: sparsity reduces cost for storage and computation. Compressed sensing [Donoho 2006, Candes & Tao 2006] If signal x is sufficiently sparse, perfect reconstruction from y = Ax with sampling rate much lower than Shannon-Nyquist bound. Regularization in classification:
Figure: Deconvolution of spike train.
2
Basis pursuit [Chen & Donoho 1999]: Given y = Ax and x unknown, x∗ = arg min x 1 , subject to y = Ax
3
The Lasso (least absolute shrinkage and selection operator) [Tibshirani 1996] x∗ = arg min y − Ax 2 , subject to x
What generates sparsity? (d’apr` es Emmanuel Cand` es) Measure first, analyze later. Curse of dimensionality.
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
(a) Harmonic functions
(b) Magnitude spectrum
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Conclusion
D.
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learification via Sparse Representation
Conclusion
This Lecture
Allen Y. Yang <yang@>
Mini Lectures in Image Processing (Part I), UC Berkeley
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Conclusion
Taking Advantage of Sparsity
1
Classification via compressed sensing Performance in face recognition Extensions
Outlier rejection Occlusion compensation
2
3
Allen Y. Yang
< yang@>
Introduction
Classification via Sparse Representation
Conclusion
Sparsity in human visual cortex [Olshausen & Field 1997, Serre & Poggio 2006]
1 2 3
Feed-forward: No iterative feedback loop. Redundancy: Average 80-200 neurons for each feature representation. Recognition: Information exchange between stages is not about individual neurons, but rather how many neurons as a group fire together.
Introduction
Classification via Sparse Representation
Conclusion
One-Pixel Camera
y = Ax, where A ∈ Rd ×D is a random projection matrix, d y1 = A(1, :)x y2 = A(2, :)x . . . yd = A(d , :)x
Introduction
Classification via Sparse Representation
Conclusion
Taking Advantage of Sparsity
What generates sparsity? (d’apr` es Emmanuel Cand` es) Measure first, analyze later. Curse of dimensionality.
Conclusion
What is Sparsity
Sparsity A signal is sparse if most of its coefficients are (approximately) zero.
(a) Harmonic functions
(b) Magnitude spectrum
Figure: 2-D DCT transform.
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Conclusion
Sparsity in spatial domain
Conclusion
Sparsity in human visual cortex [Olshausen & Field 1997, Serre & Poggio 2006]
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Conclusion
Sparsity and
Introduction
Classification via Sparse Representation
Conclusion
What is Sparsity
Sparsity A signal is sparse if most of its coefficients are (approximately) zero.
1 2
Numerical analysis: sparsity reduces cost for storage and computation. Compressed sensing [Donoho 2006, Candes & Tao 2006] If signal x is sufficiently sparse, perfect reconstruction from y = Ax with sampling rate much lower than Shannon-Nyquist bound.
1
1 -Minimization
“Black gold” age [Claerbout & Muir 1973, Taylor, Banks & McCoy 1979]
Figure: Deconvolution of spike train.
Allen Y. Yang
< yang@>
1
≤k
Allen Y. Yang
< yang@>
Compressed Sensing Meets Machine Learning
Introduction
Classification via Sparse Representation
Conclusion
Taking Advantage of Sparsity
Introduction
Classification via Sparse Representation
Conclusion
Taking Advantage of Sparsity
What generates sparsity? (d’apr` es Emmanuel Cand` es) Measure first, analyze later. Curse of dimensionality.