Research On Face Recognition Algorithm Based On SVM And Its FPGA Implementation

Face recognition technology has become one of the hot reseach topics in biometrics recognition, and it integrates computer graphics, digital image processing, computer vision, pattern recognition and artificial neural network. Face recognition technology has a broad prospect in public security and military security. Support vector machine (SVM) is a new rising machine learning method, and it overcomes several problems, such as the curse of dimensionality and the over-fitting problem. Support vector machine has been successfully used in fields, such as pattern recognition and regression analysis,so it has become a new hot research topic. So far, there are two main questions that SVM has not completely settled, the parameter selection and the application of SVM in multi-class classification.This thesis first researches on Independent Component Analysis algorithm (ICA) and Discriminative Common Vector algorithm (DCV) and applies these two algorithms to face recognition. Secondly, the fundamental principles and optimization algorithms of SVM is introduced in details. Then, this paper research on the SVM algorithms to deal with multi-class problems, such as One Against One SVM (OAOSVM), One Against All SVM(OAASVM), and Directed Acyclic Graph SVM(DAGSVM).Experiments of the classification algorithms in this paper with same feature extraction algorithm (ICA or DCV) are carried out on ORL face database using MATLAB software. And the performance comparison between SVM classifiers and minmal Euclidean distance classifier shows that SVM classifiers are better than minmal Euclidean distance classifier.After the reseach and comparison of methods, a face recognition system is realized on DE2board using Quartus II software. The system uses the face images stored in the FLASH memory on DE2board as input image, and displays the recognition results on VGA displayer. The time using different classifiers for recognizing one face image are measured. The result shows that the system using DCV+DAGSVM algorithm only takes less than40ms for recognition and the DAGSVM has better processing speed than the other two algorithms.In the end, the work of this thesis is summarized, and the future work is to complete the algorithms and accelerate the hardware system.