The Research Of Face Recognition Algorithm Based On Improved Sparse Representation

Face recognition technology is an important research field of computer vision, which has been widely used in certificate check, forensic investigation, gesture recognition, video surveillance, et al. In the past few decades, researchers have proposed many face recognition algorithms. The sparse representation based classification(SRC) has become the research focus due to its good classification performance in recent years. The related experiments have proven that SRC can obtain satisfactory results and it is not sensitive to specific feature extraction methods. But face recognition still has many challenges, such as face recognition under complex conditions caused by various illumination, posture and facial expression, et al. Besides, the small sample size problem is also a major difficulty. Therefore, how to improve the recognition rate under complex conditions and the small sample situation is a topic worth studying.This paper is focus on the sparse representation based face recognition algorithm. To solve face recognition under the complex conditions and the small sample situation, a kernel-based two-phase sparse representation method and a Gaussian weighted sparse representation based on virtual samples method have been proposed. The main research work is as follows:(1) In order to improve the classification effects of the sparse representation based face recognition algorithm, this paper introduces kernel trick to the algorithm. The face recognition is more likely to be a non-linear separable problem due to the fact that the original face image might be affected by the illumination, expression, pose, occlusion, et al. Therefore, this paper introduces kernel trick to the sparse representation based face recognition algorithm and proposes a kernel-based two-phase sparse representation method. The proposed method first exploits a non-linear function which maps the original low-dimensional data space to high-dimensional feature space. Then, we represent the test sample as a linear combination of all the training samples in the feature space and select M “nearest neighbors” of the test sample according to the representation contribution of each training sample. Finally, the test sample is represented as a linear combination of the selected M nearest neighbors and exploits the representation contribution of every class to perform classification.(2) This paper research on the solving methods of small sample size question to improve the classification accuracy rate. The related experiments show that SRC can obtain satisfactory results while providing enough training samples to represent test samples. But a real face recognition system often can obtain only a small number of training samples because of the limited storage space and the limited time of capturing images, namely the real face recognition is often a small sample size problem. In order to effectively solve face recognition under the small sample situation, a Gaussian weighted sparse representation based on virtual samples method is proposed in this paper. The proposed method first extends the training samples set by exploiting the symmetry of the face to generate virtual samples. Then, for a test sample, we exploit the Gaussian kernel distance to measure similarity relationship between the test sample and every training sample; and we put the Gaussian kernel distance as the weight of the training sample to form weighted training samples set. Finally, we exploit the sparse representation method to perform classification.(3) The test of algorithm performance. In order to test the performance of the proposed algorithms, this paper exploits ORL, FERET, AR face databases to perform experiments. The corresponding experiment results show that the proposed algorithms still can obtain better classification performance under the complex conditions and the small sample situation.