基于字典优化的稀疏表示人脸识别

Sparse Representation Based Face Recognition with Optimization of Dictionary

为消除非受控训练环境中光照/表情变化的不利影响,控制部分遮挡/伪装对人脸图像的破坏程度,提出了一种基于低秩矩阵恢复的字典优化设计,以增强稀疏表示人脸识别的性能.首先对存在非受控干扰成分的训练字典进行低秩矩阵恢复,获得相对"干净"的训练图像进行特征提取;接着采用分块相似性先验嵌入稀疏编码的方法实现对人脸图像的分类.实验结果表明,通过改进稀疏编码字典的鉴别能力,系统能更有效地抑制光照、表情、遮挡/伪装的影响,其识别的稳健性和鲁棒性得到了明显提升.

In order to eliminate the adverse impact of illumination/expression variation, and to suppress the destructiveness of facial feature contaminated by occlusion/disguise, an optimization of dictionary based on low-rank matrix recovery is investigated, which could enhance the performance of sparse representation based face recognition. At first, the training dictionary that contains the uncontrolled factor is recovered by low-rank matrix recovery, so as to obtain the ＇clean＇ training samples to finish feature extraction. Then the stage of classification is implemented by sparse coding that embeds block priori similarity. The experimental result shows that after improving the discrimination ability of sparse dictionary, not only can the system restrain the influence of illumination, expression, occlusion, but also the recognition robustness would be obviously upgraded.