The Research Of Feature Extraction In Facial Images

Human facial feature extraction is the key of face recognition. It process and analyze the input information in the model, and extract the information what is less vulnerable to interference by random factors as the feature, feature extraction has effects of increasing recognition accuracy、reducing the computation and improving the processing speed. Due to influence of illumination、facial expression、rotation gestures and other factors, there is a certain degree of difficulty in correct face recognition. Therefore, feature extraction is the main difficulty in face recognition, face feature extraction technology that emphasis on taking into account robust and efficiency has been a hot research.this article consider from the face image of the texture features, analyze and research on the local binary pattern (LBP) feature, present improvement for the limitations in LBP features. This work mainly in the following two aspects:(1)Using traditional LBP operator to describe texture information, different value of P has different detail, this article present improved LBP operator. Improved5×5LBP operator reduce24pixels of2layers to12new pixels, it contain more information than traditional3×3operator, At the same time, lower cost than the traditional5×5operator on the space and time. For the center pixel, improved5×5LBP operator has information of the4pixels around it. Through analysis and comparison of experimental results, the feature that is described by the improved operator has better classification than the traditional operator.(2) For extracting the most effective identification features of the vector, analyzing the value of each element in LBP histogram vector, this article present a method that combine LBP feature and the statistical uncorrelated optimal discriminant criterion(LBP+SUMSDC method). LBP histogram vector as a model is projected to uncorrelated projection space by optimal discriminant vectors and get new feature vector. The new feature vector has minimum separability in same class and maximum separability in different class. Through analysis and comparison of experimental results, LBP+SUMSDC method better and LBP method is effective.