Study On Face Recognition Using Scale Invariant Feature Transform And Support Vector Machine

The biological characteristics are the inherent attributes of human beings, which have strong self-stability and individual in dependency. Therefore it has been a common research problem to apply the individual biological characteristics to machine intelligence and make the computer more intelligent like human beings. Recently, several different biological characteristics have been applied in intelligent systems including face recognition, fingerprint recognition, iris recognition, palm print recognition etc. Among these, face recognition has become very popular and been improved a lot. Face recognition consists of three main parts, i.e. image preprocessing, feature extraction and pattern classification. In this paper, we mainly concentrated on the feature extraction and pattern classification parts and especially analyzed the performances of applying support vector machine into the face recognition.Image preprocessing mainly focuses on image geometry normalization, intensity histogram equalization and face area masking. These will complete the image normalization and help improve the image quality, decrease the computation complexity and therefore enhance the recognition accuracy and accelerate the convergence speed.In the image feature extraction stage, both global image features and local image features are introduced in this paper including gray-scale pixel values, Gabor Wavelet Filter, Local Binary Pattern, Scale Invariant Feature Transform, etc. In our experiments we focus more on Scale Invariant Feature Transform method and analyze the detailed performances of these different methods for representing face information.In the process of face classification, our main efforts are about using support vector machine classifier to discriminate face classes. We introduce the application in two-class and multi-class classification of support vector machines (SVM). To increase the computation speed of SVM training, we proposed a new method based on Binary Tree structure, which can also increase the robustness and accuracy.We conduct several experiments on ORL and Yale face database. The experiments consist of two main parts: one is that we compare the presentation abilities of different feature extraction methods, which demonstrate that SIFT-based method outperformed other related methods. The other one is that we compare the performance of three methods which are all based SIFT feature but use different classification strategies. The recognition results demonstrate the proposed method's robust performance under different expression conditions.