The Study On Face Image Analysis Algorithm

Computer vision and pattern recognition is one of the most important branches in computer science and artificial intelligence field. The mission of computer vision is to study about the computable model of human vision system and to construct practical vision system. This thesis presents a series of algorithms on face detection, tracking and recognition. This thesis also studies about nonlinear dimension reduction and corner point detection. The main works are as following:(1) Face detection, which aims at locating face in still images, is the pre-work of face recognition. We construct weak classifier by a Haar feature; then weak classifiers are combined to a strong classifier in a linear way. The final classifier is built in a cascade structure, which could reject most non-face samples in the early layer. Also we use integral image to quickly calculate the feature and reduce the detection time.(2) Principal component analysis (PCA) method is a popular method in face recognition field. However, PCA treats equally each feature and ignores its different role in the recognition task. A novel weighted PCA (WPCA) method is proposed for the face recognition problem. Different like classical PCA, which objects at minimizing the reconstruction error and treats equally each feature, the weighted PCA associates each feature with a coefficient according to its role in the recognition task and minimizes the weighted reconstruction error. Then classify new samples by calculating the point from weighted subspace distance for classification. The algorithm of calculating these weight coefficients by the inter-covariance and extra-covariance is also proposed. Experiment shows this weighted method can improve the recognition rate significantly compared to classical PCA.(3) Taking advantage of face detector constructed, this thesis presents a face-tracking system by optimal gradient search of confidence. The final tracking system is built on MS DirectShow platform. It runs smoothly at 15 fps for 320x280 images on P4 1.4GHz PC.This thesis also analyses and discusses about two nonlinear dimension reduction methods: LLE and Isomap. Also an angle based Isomap method is proposed. Experiments show that angle based Isomap can preserve the geometric characteristic very well. As both LLE and Isomap do not provide a transform for(4) new sample, this thesis presents a novel transform method by the neighbor relation of new input sample.(5) Corner detection has been a basic yet difficult problem in computer vision field. Here this thesis proposes a corner detection algorithm based on adaptive line approximation, which adaptively changes the length of approximation window. Then estimate the curvature by line approximation and detect corner points by curvature information.