Research On Application Of Feature Fusion And Dictionary Learning In Face Gender Recognition

Face gender recognition technique is one of the research focus in the computer vision field,which has been widely applied on the aspects of human-computer interaction, intelligent monitoring,video retrieval and so on. Like other recognition problems, feature description and classifier design is two important components of the face gender recognition system. On the one hand, face descriptive features can be divided into two major categories: global features and local features.These two kinds of features have different and complementary roles. But traditional face gender recognition methods only utilize a single feature in feature extraction, which may decrease the final recognition rate. On the other hand, traditional face gender classifiers includes SVM(Support Vector Machine), Adaboost, neural networks and so on, especially SVM is the most widely used classifier for face gender recognition because SVM is suitable for binary classification problems. However,these traditional gender classifiers perform poor for partially obstructed faces.In order to solve the above two problems, feature fusion and sparse representation are applied to face gender recognition. We fuse global and local features into a fusion feature that be taken as the face descriptive feature. Compared with a single feature, the fusion feature has more useful information of human faces. Because sparse representation has been successfully applied in the field of face recognition and has strong robustness for faces under partially obstruction and changes of lighting and expression, we apply sparse representation to face gender recognition. We further introduce the class-dictionary and dictionary learning to improve recognition accuracy. The main research contents of this thesis are summarized as follows:1. Thoroughly investigate popular face gender recognition techniques. We review commonly used methods of face feature extraction and face gender classification.2. Propose a new method for face gender recognition based on single feature and the classdictionary. Firstly, because sparse representation has been successfully applied in the field of face recognition, we propose a method for face gender recognition, which uses the global-dictionary to face gender recognition. Secondly, we replace the global-dictionary with the class-dictionary and propose a new method for gender recognition based on the class-dictionary. Lastly, we use PCA, LBP and 2D-Gabor features respectively to compare the two methods' recognition accuracy and the experiment results based on CAS-PEAL database show that the second method can achieve better performance.3. Propose a novel method for face gender recognition based on feature fusion and dictionary learning for the class-dictionary. Firstly, taking account of the different and complementary roles of global and local features, we fuse a global feature and a local feature into a fusion feature as the face descriptive feature. The experiment results show that the fusion feature can achieve better recognition rate. Secondly, because the representation ability of the untrained class-dictionary for training data is insufficient, we make use of dictionary learning algorithms to train the class-dictionary and apply the trained class-dictionary to face gender recognition. The experiment results show that the trained class-dictionary can achieve better performance in comparison with the untrained class-dictionary and RLS-DLA algorithm achieves the best recognition rate of the three kinds of dictionary learning algorithms. Then, we investigate three kinds of methods for solving sparse representation vectors. The experiment results show that the performance of OMP(Orthogonal Match Pursuit) is the best in terms of computation time and recognition rate. Lastly, based on unobstructed faces and partially obstructed faces separately, we compare the face gender recognition rate between our proposed method and SVM. The experiment results show that the performance for unobstructed faces of SVM is slightly better than our proposed method,but our proposed method has much better performance for partially obstructed faces, which proves the effectiveness of our proposed method.