有遮挡人脸识别综述:从子空间回归到深度学习

发布时间：2018-06-06 23:14

本文选题：人脸识别 + 子空间回归　；参考：《计算机学报》2018年01期

【摘要】：有遮挡人脸识别是面向现实的人脸识别系统需要重点解决的问题,其困难性主要体现在由遮挡所引发的特征损失、对准误差和局部混叠等方面.该文从鲁棒分类器的设计和鲁棒特征提取两方面回顾了现有的方法.充分利用人脸图像和遮挡自身所固有的结构来表示、抑制或消除遮挡或由遮挡引发的误差是目前设计鲁棒分类器的关键思路.从子空间回归的角度回顾了主流的线性回归分类器处理遮挡问题的一般方法:协同表示、遮挡的字典表示及遮挡字典的学习和压缩技术;从结构化误差编码的角度回顾了基于人脸图像低秩结构的误差编码方法和将遮挡的空间结构嵌入重构误差的编码方法;从噪声抑制和遮挡检测两方面回顾了现有的迭代重权误差编码方法.文中强调特征提取对于解决有遮挡人脸识别问题的重要性,总结了鲁棒特征提取的基本要素,深入分析了以图像梯度方向和韦伯脸为代表的"浅层"特征所引发的零和差异现象、以PCANet为代表的将卷积神经网络与经典的"特征图-模式图-柱状图"特征提取框架相结合的编码原理,以及以DeepID为代表的卷积神经网络所生成的"深度"特征所具有的遮挡不变性及其所蕴含的编码准则.在Extended Yale B、AR和LFW等三个基准数据库上对现有方法的有效性进行了大规模测试,指出了现有方法的适用面及局限性.最后指出了有遮挡人脸识别给计算机视觉带来的挑战、现有方法在优化算法和特征提取方面存在的主要问题以及未来利用卷积神经网络处理遮挡问题需重点考虑的问题.
[Abstract]:Face recognition with occlusion is an important problem to be solved in a realistic face recognition system. Its difficulty is mainly reflected in the feature loss caused by occlusion, alignment error and local aliasing. This paper reviews the existing methods in terms of robust classifier design and robust feature extraction. Using fully the inherent structure of face image and occlusion itself, the key idea of designing robust classifier is to suppress or eliminate the occlusion or the error caused by occlusion. From the point of view of subspace regression, the general methods of mainstream linear regression classifier to deal with occlusion are reviewed: collaborative representation, dictionary representation of occlusion and learning and compression techniques of occlusion dictionary; From the point of view of structured error coding, the error coding method based on low rank structure of face image and the coding method of embedding occluded spatial structure into reconstruction error are reviewed. The existing iterative weighted error coding methods are reviewed from the aspects of noise suppression and occlusion detection. This paper emphasizes the importance of feature extraction in solving the problem of shaded face recognition, and summarizes the basic elements of robust feature extraction. The zero-sum difference caused by the "shallow" features, represented by the gradient direction of the image and the Weber face, is analyzed in depth. The coding principle of combining convolutional neural network with the classical feature extraction framework of "feature graph-pattern graph-bar graph", represented by PCANet, is proposed. The occlusion invariance of the "depth" feature generated by the convolution neural network represented by DeepID and its encoding criterion are also discussed. The validity of the existing methods is tested on the basis of three benchmark databases, Extended Yale Bearar and LFW, and the applicability and limitation of the existing methods are pointed out. Finally, it points out the challenge to computer vision caused by face recognition with occlusion, the main problems existing in the existing methods in optimization algorithm and feature extraction, and the problems that need to be considered in the future using convolution neural network to deal with occlusion problem.
【作者单位】：浙江工业大学计算机科学与技术学院;
【基金】：国家自然科学基金(61402411,61379017,61672464) 浙江省自然科学基金(LY18F020031,LY17F020021)资助~~
【分类号】：TP181;TP391.41

【相似文献】