Research On Facial Feature Localization Based On Support Vector Machine

Face recognition and analysis have been the research focuses in the area of the machine vision and artificial intelligence. As an important step of face recognition and the foundation of face analysis, facial feature localization has a significant impact on the reliability of face recognition and analysis.After investigated a variety of algorithms within the field of facial feature alignment and in-depth study of the Active Shape Model(ASM), a Support Vector Machine(SVM)-based facial feature localization algorithm is proposed in this thesis.Aiming at the shortages of ASM, such as sensitivity to initial face shape and ambiguity of local feature’s texture,three improvements have been made:1. SVM classifier based on Histograms of Oriented Gradients(HOG) descriptor is used to iden-tify the local feature,and the posterior of SVM is used to estimate the similarity between one pixel and local feature.Better description of local feature and estimation of similarity are got with this algorithm.2. The probability of face shape similarity is proposed to evaluate the overall face shape based on the SVM posterior, it is used to guide the position update of local feature and the judge-ment of face shape convergence to guarantee the convergence of iteration.3. A method to estimate the initial face shape via part of the local features is proposed. The part of local features are randomly selected according to their posterior and the probability of face shape similarity.The method has been tested on the BioID database, the results show that the method proposed is robust to small-angle deflection,expression change and occlusion of part of local features.A detection rate around90%is achieved.