欠完备采样环境下面向数据的稀疏表示人脸识别研究

发布时间:2019-06-22 20:33
【摘要】:人脸识别因其非接触性、易采集等优点而被广泛研究,并主要应用于考勤、门禁、监控、公安等系统。虽然目前众多种人脸识别算法已取得较好的识别性能,但人脸识别系统在实际应用中仍面临很多挑战,其中主要包括:由光照变化、饰品等引起的遮挡人脸识别问题;非可控条件下可采集到的样本数少,即小样本问题;姿势和表情变换等,本文将这种情况下的人脸识别称之为欠完备采样人脸识别。欠完备采样会造成人脸信息的缺失,降低已有算法的识别性能。为此提出面向数据的稀疏表示算法对该问题展开研究以提高人脸识别算法的鲁棒性和实用性,具体研究内容如下:(1)根据近邻表示和分辨性分解算法分别提出基于近邻类加权结构稀疏表示图像识别算法和基于分辨性分解结构稀疏表示遮挡人脸识别算法。字典中各类训练样本对测试样本分类的贡献不同,一般近邻样本对测试样本的正确分类具有较大的促进作用,因此考虑选取近邻类并加权进行测试样本分类,不仅可降低算法的计算复杂度,同时提高算法识别率。此外,为提高遮挡情况下的人脸识别性能,采用分辨性分解算法对遮挡部位进行分离,并在分解得到的共同部分和低秩条件部分上分别进行主成分分析,并计算投影矩阵,最后在投影空间上进行结构稀疏表示并分类。(2)为解决全局算法对遮挡的敏感性,并进一步降低遮挡对识别性能的影响,对图像进行分块局部处理,通过对遮挡模块赋予低权值,干净模块赋予高权值来降低遮挡模块对算法性能的影响。为此,本文提出几种不同的模块加权方案:首先,将图像分割成多个有重叠的模块,并利用Fisher率计算每个模块的分辨性,据此对每个模块加权,保留分辨性高的模块进行后续的分类识别;其次,将图像分割成4部分,并利用稀疏残差对模块加权进而对遮挡部分进行估计,最后仅在非遮挡部位上进行分类判别;最后,将以上两种加权方案联合,提出基于Fisher判别和稀疏残差的模块加权算法,该算法联合了Fisher加权和残差加权的优势,以进一步提高遮挡检测性能。(3)为精确检测遮挡部位,并实现无遮挡训练集上的遮挡人脸识别,提出两种像素级上的遮挡检测算法:基于稀疏表示的像素级遮挡检测人脸识别及基于块递推残差分析的双层稀疏表示分类算法。基于稀疏表示的像素级遮挡检测算法根据类残差分析各类遮挡估计结果,并对结果进行统计得出最终的像素遮挡估计,最后仅在非遮挡像素集上进行识别。基于块递推残差分析的算法将遮挡样本分成上下两个模块,利用稀疏度较高的模块重构整幅测试图像,并根据残差估计遮挡像素进而对各像素进行加权并分类,以此提高遮挡人脸的识别性能。像素级遮挡检测可避免分块遮挡检测算法中模块中同时含遮挡和非遮挡部分而造成的识别率低的问题。(4)利用核空间对块稀疏表示算法进行非线性扩展,并提出核块稀疏表示算法(KBSRC:Kernel Block Sparse Representation based Classification),该算法将样本投影到降维的核空间,因而可将样本的原非线性空间线性化,而在该空间上利用样本的结构信息分类可提高分类性能。
[Abstract]:Face recognition is widely studied because of its non-contact and easy acquisition. It is mainly applied to the system of attendance, entrance guard, monitoring and public security. Although a large number of face recognition algorithms have acquired better recognition performance, the face recognition system still faces many challenges in the practical application, which mainly comprises the problems of shielding face recognition caused by illumination changes, ornaments, and the like; and the number of samples that can be acquired under the non-controllable condition is small, In this paper, the face recognition in this case is called the under-complete sampling face recognition. Under-complete sampling can cause the loss of face information and reduce the recognition performance of the existing algorithm. To this end, a sparse representation algorithm for data is proposed to study the problem to improve the robustness and practicability of the face recognition algorithm. (1) based on the nearest neighbor representation and the resolution decomposition algorithm, a sparse representation image recognition algorithm based on a neighbor class weighting structure and a sparse representation shielding face recognition algorithm based on the resolution decomposition structure are respectively proposed. The contribution of various training samples in the dictionary to the classification of the test samples is different, and the common neighbor samples have a great effect on the correct classification of the test samples, therefore, considering the selection of the nearest neighbor class and weighting the test sample classification, not only can the computational complexity of the algorithm be reduced, Improve that recognition rate of the algorithm at the same time. In addition, in order to improve the face recognition performance in the case of occlusion, the shielding part is separated by a resolution decomposition algorithm, and the main component analysis is carried out on the common part and the low-rank condition part obtained by the decomposition, and the projection matrix is calculated, And finally, structural sparse representation and classification are carried out on the projection space. (2) in order to solve the sensitivity of the global algorithm to the occlusion, and further reduce the influence of the occlusion on the recognition performance, the image is segmented and partially processed, and the high-weight value is given by the clean module to reduce the influence of the blocking module on the performance of the algorithm by giving a low weight to the shielding module. To this end, several different module weighting schemes are proposed: first, the image is divided into a plurality of modules with overlapping, and the resolution of each module is calculated by using the Fisher rate, the method comprises the following steps of: dividing an image into four parts, weighting the module by using a sparse residual error to estimate the shielding part, and finally carrying out classification judgment on the non-shielding part; and finally, combining the two weighting schemes, and putting forward a module weighting algorithm based on the Fisher discrimination and the sparse residual, The algorithm combines the advantages of Fisher's weight and residual weight to further improve the shielding performance. And (3) in order to accurately detect the occlusion region and realize the occlusion face recognition on the non-occlusion training set, the occlusion detection algorithm on the two pixel levels is proposed, namely, the pixel-level occlusion detection face recognition based on the sparse representation and the double-layer sparse representation classification algorithm based on the block recursive residual analysis. The pixel-level occlusion detection algorithm based on the sparse representation analyzes the various occlusion estimation results according to the class residual, and then counts the results to obtain the final pixel occlusion estimation, and finally, the pixel-level occlusion detection algorithm is only identified on the non-occlusion pixel set. Based on the algorithm of the block recursive residual analysis, the occlusion sample is divided into upper and lower modules, the whole image is reconstructed by using a module with higher sparsity, and the occlusion pixel is estimated to be weighted and classified according to the residual estimated occlusion pixel so as to improve the identification performance of the occlusion face. The pixel-level occlusion detection can avoid the problem of low recognition rate caused by the blocking and non-blocking part in the module in the block-blocking detection algorithm. (4) The kernel space is used for non-linear expansion of the block sparse representation algorithm, and a kernel block sparse representation algorithm (KBSRC: Kernel Block Sparse Representation based Classification) is proposed, and the sample is projected into the reduced-dimension nuclear space, so that the original non-linear space of the sample can be linearized, And the classification performance can be improved by using the structure information classification of the samples in the space.
【学位授予单位】:燕山大学
【学位级别】:博士
【学位授予年份】:2016
【分类号】:TP391.41

【相似文献】

相关期刊论文 前10条

1 李映;张艳宁;许星;;基于信号稀疏表示的形态成分分析:进展和展望[J];电子学报;2009年01期

2 赵瑞珍;王飞;罗阿理;张彦霞;;基于稀疏表示的谱线自动提取方法[J];光谱学与光谱分析;2009年07期

3 杨蜀秦;宁纪锋;何东健;;基于稀疏表示的大米品种识别[J];农业工程学报;2011年03期

4 史加荣;杨威;魏宗田;;基于非负稀疏表示的人脸识别[J];计算机工程与设计;2012年05期

5 高志荣;熊承义;笪邦友;;改进的基于残差加权的稀疏表示人脸识别[J];中南民族大学学报(自然科学版);2012年03期

6 朱杰;杨万扣;唐振民;;基于字典学习的核稀疏表示人脸识别方法[J];模式识别与人工智能;2012年05期

7 耿耀君;张军英;袁细国;;一种基于稀疏表示系数的特征相关性测度[J];模式识别与人工智能;2013年01期

8 张疆勤;廖海斌;李原;;基于因子分析与稀疏表示的多姿态人脸识别[J];计算机工程与应用;2013年05期

9 李正周;王会改;刘梅;丁浩;金钢;;基于形态成分稀疏表示的红外小弱目标检测[J];弹箭与制导学报;2013年04期

10 胡正平;赵淑欢;李静;;基于块稀疏递推残差分析的稀疏表示遮挡鲁棒识别算法研究[J];模式识别与人工智能;2014年01期

相关会议论文 前3条

1 何爱香;刘玉春;魏广芬;;基于稀疏表示的煤矸界面识别研究[A];虚拟运营与云计算——第十八届全国青年通信学术年会论文集(上册)[C];2013年

2 樊亚翔;孙浩;周石琳;邹焕新;;基于元样本稀疏表示的多视角目标识别[A];2013年中国智能自动化学术会议论文集(第五分册)[C];2013年

3 葛凤翔;任岁玲;郭鑫;郭良浩;孙波;;微弱信号处理及其研究进展[A];中国声学学会水声学分会2013年全国水声学学术会议论文集[C];2013年

相关博士学位论文 前10条

1 李进明;基于稀疏表示的图像超分辨率重建方法研究[D];重庆大学;2015年

2 王亚宁;基于信号稀疏表示的电机故障诊断研究[D];河北工业大学;2014年

3 姚明海;视频异常事件检测与认证方法研究[D];东北师范大学;2015年

4 黄国华;蛋白质翻译后修饰位点与药物适应症预测方法研究[D];上海大学;2015年

5 王瑾;基于稀疏表示的数据收集、复原与压缩研究[D];北京工业大学;2015年

6 王文卿;基于融合框架与稀疏表示的遥感影像锐化[D];西安电子科技大学;2015年

7 解虎;高维小样本阵列自适应信号处理方法研究[D];西安电子科技大学;2015年

8 秦振涛;基于稀疏表示及字典学习遥感图像处理关键技术研究[D];成都理工大学;2015年

9 薛明;基于稀疏表示的在线目标跟踪研究[D];上海交通大学;2014年

10 孙乐;空谱联合先验的高光谱图像解混与分类方法[D];南京理工大学;2014年

相关硕士学位论文 前10条

1 吴丽璇;基于稀疏表示的微聚焦X射线图像去噪方法[D];华南理工大学;2015年

2 赵孝磊;基于图像分块稀疏表示的人脸识别算法研究[D];南京信息工程大学;2015年

3 黄志明;基于辨别式稀疏字典学习的视觉追踪算法研究[D];华南理工大学;2015年

4 张铃华;非约束环境下的稀疏表示人脸识别算法研究[D];南京信息工程大学;2015年

5 贺妍斐;基于稀疏表示与自适应倒易晶胞的遥感图像复原方法研究[D];南京信息工程大学;2015年

6 杨烁;电能质量扰动信号的稀疏表示/压缩采样研究[D];西南交通大学;2015年

7 应艳丽;基于低秩稀疏表示的目标跟踪算法研究[D];西南交通大学;2015年

8 梁晓捷;基于网络摄像头与稀疏表示分类法的实时人脸识别系统应用研究[D];五邑大学;2015年

9 张宏乐;语音信号稀疏表示方法研究[D];太原理工大学;2016年

10 郭欣;基于K-SVD稀疏表示的语音增强算法研究[D];太原理工大学;2016年



本文编号:2504931

资料下载
论文发表

本文链接:https://www.wllwen.com/shoufeilunwen/xxkjbs/2504931.html


Copyright(c)文论论文网All Rights Reserved | 网站地图 |

版权申明:资料由用户0d340***提供,本站仅收录摘要或目录,作者需要删除请E-mail邮箱bigeng88@qq.com