二次近邻稀疏重构法及人脸识别

Sparse reconstruction algorithm based on secondary nearest neighbor and face recognition

基于整个数据集的稀疏表示(sparse representation classification,SRC)用于人脸识别在很大程度上影响了运行效率。如何利用较少样本稀疏表示在保证计算效率的同时,识别率也有一定提升,尤其是面对光照、角度、姿态等非受控环境,目前仍是一个问题。考虑到协同表示(collaborative representation classification,CRC)基于l2范数稀疏求解的优势,为进一步提升CRC的整体分类性能,引入类内近邻,提出一种二次近邻稀疏重构表示法。该方法首先在原始训练集上选择各类训练样本中与待测样本距离相近的若干样本组成近邻样本集,并协同表示,接着分别用各类近邻样本重构待测样本,再次选择与待测样本相近的若干重构样本协同表示,最终实现模式分类。在ORL和FERET数据库上的仿真实验表明,相比现有的一些CRC算法,该方法在一定程度上缩短了运行时间,并使识别更精确。

Sparse representation classification(SRC) based on the entire data set for face recognition largely affect the running efficiency.How to use the few samples for sparse representation while ensuring the computing efficiency,the recognition rate also has a certain improvement,especially in the light,angle,attitude and other uncontrolled environment,it is still a problem.Taking into account the advantage of sparse solution based on l2 norm in collaborative representation classification(CRC),on this basis,in order to further improve the overall classification performance of CRC,this article introduces the nearest neighbor of the inner class,a sparse reconstruction method based on secondary nearest neighbor is proposed.Firstly among the original training sample set,several samples of the inner class which are similar to the testing sample were chosen to construct the nearest neighbor sample set,and they collaboratively represent the testing sample,and the nearest neighbor samples in each class were used to reconstruct testing sample respectively,then some reconstructed samples which are similar to the testing sample were chosen to collaboratively represent again,finally pattern classification was realized.Experiments on the ORL and FERET database indicate that compared with some exitsing CRC algorithms,the proposed method partly makes the running time short,and the recognition rate more accurate.