Abstract: Single sample per person makes face recognition much more difficult. According to the fact that facial variations in same category should have similar coding coefficients, the proposed variance dictionary learning method with structured sparsity can effectively represent facial variance. Considering that all local regions from same person have same class label, query image is represented by gallery images patches and variance dictionary patches. Structured sparsity constraints are imposed on the reconstruction coefficients to automatically select corresponding class dictionary so that query image can be well represented. The proposed method can harvest the advantage of both local methods and holistic methods, and performs well compared with the existing solutions to the single sample problem on the AR, Extended Yale B, CMU-PIE datasets.

Key words: single sample per person, structured sparsity, intra-class variant dictionary, joint representation

摘要： 针对单样本问题，基于相同类别的人脸变化信息应有相似的稀疏编码这一事实，提出结构化稀疏变化字典学习方法，以得到较好的共享类内变化字典。同时鉴于同一人脸的所有区域应有相同的类标签，通过训练样本与变化字典按坐标分块联合表示查询人脸区域，然后给稀疏系数引入导致结构化稀疏效果的约束条件，实现对应类别字典的自动选择，从而更好地表示查询人脸。提出的人脸表示方法可以在局部识别方法的优势上整合全局信息，使得在AR、Extended Yale B、CMU-PIE人脸库上的表现超过其他单样本识别相关的方法，取得了较好的识别效果。

关键词: 单样本, 结构化稀疏, 类内变化字典, 联合表示