Feature Extraction Based On LBP And Deep Learning

In the information time, how to recognize and identify a person accurately and to protect information security has been become a key social problem that must to be solved. As one of the most development potential biometric identification technology, face recognition technology has got very wide application prospects with its unique advantages such as hidden operability, non-direct contact, interaction, etc. Also, the stand or fall of feature extraction has a crucial influence to the result of face recognition. Therefore, how to extract the stable and effective facial features that contains class information as much as possible to recognition, and how to combining a variety of different features to achieve more ideal classification results are all the research hotspots of current face recognition.With reading a lot of domestic and foreign relevant literatures, this paper summarizes the previous research results and studies the feature extraction methods based on the local binary pattern and deep learning. The main research work includes:(1)It is not enough that capture various facial information using only one descriptor to face features extraction. Though, DCT(Discrete Cosine Transform) can extract the frequency feature of face image, it ignores the relationships between the adjacent pixels and abandoned the texture information. ELBP(Elongated Local Binary Pattern) considered the direction in local area and the texture information but without global information. In this paper, a novel method was proposed that addressed the problems by fusing DCT and ELBP. The centralized DCT coefficients are used as frequency feature, and local features of the mouth and eyes area are extracted by ELBP. Then, fusing the two features by PCA to get more effective features. Finally, we chose 1-NN classifier to evaluate the proposed feature. Experiments on ORL and Yale face database show that the proposed method is better than just single DCT、ELBP method or the fusion method of LBP and DCT, and also improves the accuracy of face recognition.(2) Deep Belief Networks(DBNs) method ignores the image local structure and is difficult to learn the local characteristics of face image, the network training time is also too long to make full use of the facial texture feature and reduce the bad influence of illumination, a novel method by fusing local binary pattern(LBP) and DBNs features is proposed in this paper. In this method, the LBP feature is as the input of the DBNs. Furthermore, in order to accelerate the training speed of DBNs, the Extreme Learning Machine(ELM) is introduced into the network training process. Finally, the training network is used to classify and recognize. Experiments on ORL and FERET face databases with different resolution demonstrate that the proposed method is better than other relevant methods.