Research On Algorithm Of Face Recognition With Single Sample Per Person

As an important branch of biometric recognition, in recent years, face recognitionattracted more and more researchers’ attention and achieved rapid development. In somepractical applications, each person can only get one picture as the training sample. Butmost of the face recognition techniques will suffer serious performance drop when there isonly one training sample per person. As the recognition performance is not so ideal whenthere is only one sample, we do some research on the algorithm of face recognition withsingle sample per person. The main content of this paper is as follows:1. The background and development of the face recognition both at home and abroad areintroduced in this paper. Then the algorithms of face recognition with single sample perperson are summarized, the advantage and disadvantage of these methods are alsoanalyzed.2. A face recognition algorithm based on sample augment methods and improved twodimensional principal component analysis (2DPCA) is proposed. Some of the sampleaugment methods are combined to synthesize virtual samples in order to make full use ofthe single training image. Improved2DPCA is chosen to extract the feature of the syntheticvirtual samples. The training samples are divided into sub-blocks and then the covariancematrix is constructed by these sub-blocks which are normalized by the within-class averagevalue of each sub-block. The proposed algorithm is better than general2DPCA and betterthan using only one sample augment method.3. Face recognition techniques will suffer serious performance drop when there is onlyone training sample per person, especially when the face is under varying lightingconditions. Extracting illumination invariant features through nonsubsampled contourlettransform can improve the recognition rate of the single sample face recognition underillumination variations. In order to eliminate the impact of the partial shadow on the final recognition, a serial manner of local matching strategy is proposed in this paper. Firstly, thewhole image is used for coarse classification, and then a decision is made whether it need anext layer of classification which combine the classification results of the global featureand each sub-block. The proposed method of local matching can effectively solve theproblem of single sample face recognition under varying lighting condition. This methodreaches a balance between recognition accuracy and speed. It not only increases therecognition accuracy but also improves the speed of classification.