基于WSSRC单样本人脸识别及样本扩充方法研究

Single-sample Face Recognition Based on WSSRC and Expanding Sample

摘要:

由于传统的SRC方法的实时性不强、单样本条件下算法性能低等缺点,提出了融合全局和局部特征的加权超级稀疏表示人脸识别方法(WSSRC),同时采用一种三层级联的虚拟样本产生方法获取冗余样本,将生成的多种表情和多种姿态的新样本当成训练样本,运用WSSRC算法进行人脸识别分类。在单样本的情况下,实验证实在ORL人脸库上该方法比传统的SRC方法提高了15.53%的识别率,使用在FERET 人脸库上则提高7.67%。这样的方法与RSRC 、SSRC、DMMA、DCT-based DMMA、I-DMMA相比,一样具备较好的识别性能。

Abstract:

Due to the shortcomings of the traditional SRC in face recognition, a face recognition method with one training image per person has been proposed, and it is based on compressed sensing. We apply three level cascades virtual sample method to generate multiple samples of each person. These generated samples have multi-expressions and multi-gestures are added to the original sample set for training. Then, a super sparse random projection and weighted optimization are applied to improve the SRC. This proposed method is named weighted super sparse representation classification (WSSRC) and is used for face recognition in this paper. In the case of the single sample,,experiments on the well-known ORL face dataset show that WSSRC is about 15.53% more accurate than the original SRC method and on the FERET face dataset,it is increased by 7.67%.In addition, compared to RSRC ,SSRC,DMMA,DCT-based DMMA and I-DMMA, WSSRC also achieve higher recognition rates .