基于随机投影与加权稀疏表示残差的光照鲁棒人脸识别方法

Illumination robust face recognition based on random projection and weighted residual of sparse representation based classification

针对人脸识别中的光照变化问题,利用随机投影对传统稀疏表示分类器进行改进,提出一种基于随机投影与加权稀疏表示残差的光照鲁棒人脸识别方法。通过对人脸图像进行光照规范化处理,尽量消除人脸图像上的恶劣光照,取得经光照校正的人脸样本后进行多次随机空间投影,进一步丰富样本的光照不变特征,以减小光照变化对人脸识别带来的影响。在此基础上,对利用单一残差分类的传统稀疏表示分类方法进行改进,样本经过多次随机投影和稀疏表示会产生多个样本特征和重构残差,利用样本特征的能量来确定各个重构残差的融合权值,最终得到一种稳定性和可靠性更强的加权残差。在Yale B和CMU PIE两个光照变化较大的人脸库上的实验结果表明,改进的方法具有较强的光照鲁棒性。与传统稀疏表示方法相比,本文提出的方法在Yale B人脸库上两组实验的平均识别率分别提高了25.76%和46.39%,在CMU PIE上的平均识别率提高了10%左右。

To solve the problem of illumination variation in face recognition, we improve the tradi tional sparse representation classification （SRC） by using random projection, and propose an illumina tion robust face recognition method based on random projection and weighted residual of sparse represen ration based classification （RPWR SRC）. By normalizing the illumination of the face image, the bad illu mination on the face image is eliminated as much as possible. Then we introduce random projection into the proposed method. The face images with illumination correction are projected to multiple random spaces to enrich the illumination invariant features of the samples and to reduce the influence of illumina tion changes on face recognition. On this basis, we utilize the single residual classification to improve the traditional sparse representation classification method. By multiple random projections and sparse repre sentations, we can obtain multiple sample features and reconstructed residuals. With the energy of sam pie features, we determine the weights for fusion of respective reconstructed residual to get a more sta ble and reliable weighted residual instead of single residual in traditional SRC. Experiment results on Yale B and CMU PIE face databases show that, the proposed RPWR SRC has strong robustness for face recognition under different illumination conditions. Compared with the traditional SRC, its average rec ognition rates of the two experiments on Yale B face database are increased by 25.76% and 46.39% re spectively, and the average recognition rate on CMU PIE is increased by about 10%.