Design And Implementation Of Face Recognition Algorithm Based On Sparse Representation

As an important sensor nodes of ubiquitous network, face recognition is widely used in many areas such as authentication and video surveillance. The application scenarios mentioned above have strict limitation for the image resolution and data storage space, so we need a more efficient algorithm. The face recognition algorithm based on sparse representation classification has achieved higher performance on several public face databases in laboratory experiments. But in real-world application, it do not achieve a same result, because of the environment factors. Based on the analysis of the disadvantage of sparse representation classification, we proposed a sparsity concentration discriminant analysis algorithm.Based on sparse representation classification, we have proposed an sparsity concentration discriminant analysis algorithm, called SCDA. Firstly, we have designed a sparse concentration index as an evaluation criterion for each sample., and pick up the poor test sample. Then we choose a properly threshold to screen the test samples and modify the training base. Finally, we classify again based on the simplified training base. To improve the recognition rate, we proposed a multi-layer structure for face recognition by filtrating several times. We have designed several experiments to compared the effectiveness of this algorithm on the ORL, Extended Yale B and UMIST face databases. In the test on the face databases, our method has a good adaptability that it can get 90% accuracy without noise. And it can get more than 70%with the 50% polluted sample. Compare to sparse representation classification, the sparsity concentration discriminant analysis improve the accuracy by 5%-30% with only 2% entire test time increased.Based on the above research work, we have developed a real-time face recognition system with the help of MFC and OpenCV library. After that, we test the system in the real-world scenes, and the accuracy reached 83.33%. After deeply tested about the algorithm in the personal computer, we transplanted the system to the SEP0611 chip, and optimized according to the characteristics of the embedded platform. The accuracy could reach 75% after test.