Abstract: Subspace Pursuit（SP） algorithm needs prior knowledge of sparseness in face recognition based on sparse representation because its candidate atoms have the same number with sparseness. Against the disadvantage, this paper proposes an improved subspace pursuit algorithm. This algorithm introduces backtracking iterative optimization algorithm in atom selection to ensure the candidate atoms increase with the number of iterations simultaneously. By removing the candidate atoms with low credibility to make sure that the chosen atoms have the most similar structure with the identifying face image, to ensure that the presented algorithm can reconstruct faces well. The experimental results show that the improved algorithm have the highest recognition rate in face recognition on ORL and Yale B face database respectively compared with SP, Sparsity Adaptive Subspace Pursuit（SASP）, Orthogonal Matching Pursuit（OMP） and the OMP-cholesky algorithm employing in the Sparse Representation Classification（SRC） frame.

Key words: sparse coding, sparse representation, face recognition, Orthogonal Matching Pursuit, subspace tracking

摘要： 基于稀疏表示的人脸识别中的子空间追踪（SP）算法的候选原子个数固定与稀疏度相同，因此需要已知信号的稀疏度。针对该缺点，提出一种改进的子空间追踪算法，在选择原子的过程中引入回溯迭代优化思想，候选原子个数随着迭代次数逐一增加。通过移除候选原子集中数量同样逐一增加的可信度较低的原子，使选择的原子与待识别人脸图像具有最相似的结构，能较好地重构人脸。采用稀疏表示分类（SRC）框架，分别与基于SP、SASP、正交匹配追踪（OMP）、OMP-cholesky的人脸识别相比，在ORL和Yale B人脸数据库上的实验结果表明，该算法有最高的识别率。

关键词: 稀疏编码, 稀疏表示, 人脸识别, 正交匹配追踪, 子空间追踪