Face Recognition Based On Scale Invariant Features

As a branch of biometrics, human face research is becoming a hot international focus in recent years. However, face recognition research faces many problems, such as face location in image, face normalization, features extraction and some more challenging problem such as complex Illumination and pose. While all these research will based on fit features, this paper is committed on finding better ways using existing features and exploring new features in face recognition field. The main contribution is as follows:1. Face recognition under visible light is a difficult problem, face normalization is necessary if effective face recognition is available under visible light, while eye location is the key problem. This paper proposes a new eye location method using eigenfeature and Gabor feature template. By idea of step location, candidate eye positions are located firstly, then these candidate positions are filtered, finally the eye position is determined. Different features are used in different location step to play advantage of their own characteristics in finishing eye location problem under complex illumination collaboratively. 2. Gabor features are very effective in face recognition, which have strong ability in expressing image feature due to multi-resolution characteristic, while Gabor features are limited due to high dimension at the same time. This paper proposed a Gabor feature selection method. Gabor features are filtered during extraction. This feature reducing method makes it convenient in using and calculating Gabor feature, keeping the classification performance of Gabor feature meanwhile. The experiment shows its superior performance compared to similar algorithms.3. Speed-Up Robust Features (SURF) is a scale and in-plane rotation invariant feanture with comparable or even better performance with SIFT. Because SURF feature is the local feature of the interest points, how to use these feature vectors to recognize face is the basic problem. We proposed a similarity function to evaluate the similarity of two face images using the result information of the point match. This made us to exploit SURF features in face recognition successful. Compare with SIFT features, SURF have obviously lower time consumption and similar recognition performance..4. Multi-pose recognition is a challenge problem in face recognition. We hope use few training samples to achieve higher recognition performance. In this thesis, the SIFT feanture is combine with the feature pool to build the correlated features,and change the problem from recognition different pose image to classifier the correlated features. To classifier the correlated correctly, we propose the SVM classifier of sub-space classifier and SVM full-space classifier. A features select method is used to uniform the feature dimension of the feature pool, so that the SVM full-space classifier can be used.