Research On Face Recognition Based On Sparse Representation And Artificial Neural Network

The technology of face recognition is indispensable to our modern society. The essence of the identity authentication technology is a problem of intelligent classification, it is a concrete manifestation of artificial intelligence. It is an important biometric technology that extracts visual characteristic information from facial images toachieve identification and has very broad prospects for development. Face recognition is a secretive non-contact operation with uniqueness and reliability,so it can be widely applied in many fields such as video surveillance and ID authentication.With high rapid development of the economy, the technology develop to the more efficient, stable and safe direction. So the subject of research still has important research value and significance. Face recognition technology involves many disciplines, on the mathematical basis of high demand, digital image processing, pattern recognition, artificial neural networks, artificial intelligence and machine learning and other subjects are covered. The technology is the current research focus and has great challenges, such as anti-jamming capability is not strong, feature extraction results are poor, recognition rate and time to be to continue to improve, and so on.There are many algorithms for face recognition technology at present, such as template matching method, the smallest neighbor method, neural network and sparse expression method and so on. These methods all have their own advantages, also has its own disadvantages, they are gray space as the expression of face image image source. Grayscale image contain a lot of noise and redundant information. Performance would be poor if we get classified information directly by this. Sparse transform coding algorithm can expressed facial feature information well. The algorithm can eliminate noise and redundant information effectively, it reserves feature information for classification. LDA algorithm can reduce the dimension of data effectively, it finds the best direction and dimension classification of coordinates, and reduces the amount of calculation. Classifier is the final step to bear face recognition classification. Artificial neural network has a broad and successful application in pattern recognition. This paper proposes a joint method of sparse expression and improved LDA to extract features, and then use the RBF network to do classify sample. Results show that the method has good recognition performance.1) KSVD algorithm is used to optimize the dictionary. The sparse coding can be got according to the dictionary and then mapped to a low-dimensional subspace with more favorable classification by LDA algorithm.2) Basis on the above algorithm, we rewrite the between-class scatter matrix in LDA algorithm. It improved sparse coding in the low-dimensional subspace which projected onto the unit sphere that avoids the near-coded overlay from the center of the sphere, Simulation results show that the algorithm improves the recognition rate.3) The sparse coding was classified by optimized RBF neural network and SVM. According to mentor’s project, we set up a uygur face database that contain 253 people, and the database was compared with the international common database according to the algorithm.Compared with other algorithm,the results show that the algorithm based on sparse expression and neural network is superior to other algorithms which with high recognition rate and fast recognition speed, but the training time is longer than the others. In addition, the algorithm has strong robustness and generalization ability to other algorithm.