|
The face recognition technology is based on the person's facial features of a biometric identification technology, face recognition complex and have a wide range in the field of image processing, pattern recognition, physiology, psychology, computer vision application, so that it has a very strong scientific research challenging. International researchers study hotspot. The key steps of the Face Recognition feature extraction and classification. This paper focuses on the feature extraction and classification recognition part. The paper first analyzes the background and development of face recognition, an overview of the steps of face recognition, face recognition performance evaluation criteria is summarized in several commonly used methods for face recognition and analysis of the major advantages of these methods disadvantages, draw will combine a variety of methods, is conducive to the conclusion of the recognition rate improved. Secondly, the wavelet transform theory, as well as face recognition, based on principal component analysis (PCA) feature extraction method and based on non-negative matrix factorization (NMF) feature extraction method, and the latter two method for comparison. Considering the superior performance of the NMF algorithm feature extraction, this paper as feature extraction algorithm based on wavelet and NMF algorithm. And then to study the structure of the BP neural network learning methods, standard BP algorithm, BP network training steps. The network is very easy to construct, and input data are no special requirements, theoretical studies have been very mature, and widely used in practice, so this article using BP neural network as a classifier to identify. On this basis, a NMF-based feature extraction and face recognition BP neural network, and according to the method of experimental design process. Training samples and the number of test samples, determined face database selection using db2 wavelet basis and NMF decomposition, image feature extraction, using the training function BP networks traingdm training, complete training training samples and Test sample identification. After verified the effectiveness of the proposed algorithm, this experiment is completed on ORL face database, respectively, by heads in a different dimension, a different hidden layers, different learning rate and sparsity The recognition rate of the test samples were compared, and then select the highest recognition rate of the number of dimensions under method compared with other methods. Experimental results show that the proposed method in the recognition rate and the stability were improved to some extent. Finally, review and summarize the work done, the outlook for future work.
|