Semi-supervised Mixed Sparse Representation based Classification for Face Recognition

Abstract

The purpose of this paper is to solve the problem that the local optimum is prone to arise in Semi-Supervised Sparse Representation based Classification (S3RC), so as to improve the classification effect. Although S3RC can solve the issue of face recognition in the case of insufficient samples and corruption (including linear and non-linear) by using a Gaussian Mixture Model, due to the defect of Expectation-Maximization (EM) algorithm, which can only get local extremum but not global extremum, prototype dictionary (which contains only class-specific information) has a great contribution to the final recognition effect. This paper introduces a Semi-supervised Mixed Sparse Representation (SMSR) method based on dictionary decomposition to construct a prototype dictionary to solve this problem. We decompose the training image multiple times to ensure a more suitable prototype dictionary on the basis of reducing the corruption of the training data. The experiments results on AR and FERET databases demonstrate the effectiveness that, the proposed approach yields improved results compared to S3RC and other state-of-the-art algorithms.