Research On Face Verification Across Aging Based On Feature Learning

In computer vision,face verification is an important aspect of biometrics,and also a hot spot.It has a wide application prospect in many fields,such as records management system,authentication system,credit card verification,criminal identification system of public security,banking and customs monitoring,human-computer interaction,and so on.In order to address the problem caused by age-related changes and feature representation,we focus on several face verification methods combined with deep learning and attribute prediction.The main work of this paper is summarized as follows:1)The traditional hand-crafted features are not sufficient to characterize the face,we propose the method of face verification across aging based on automatic learned features and artificial features.Firstly,this method uses a nine-layer deep convolutional neural network to automatically learn rich abstract features,then in order to take full advantage of the abstractness of learned deep features and strong targeted representation of manual features,we concatenate the features extracted by fully-connected layer and hand-crafted LBP features to form final face feature vectors and use cosine similarity to measure distances between these features for face discrimination.Experimental results on CACD and LFW datasets show that the proposed method is superior to other methods in face verification accuracy.2)In unconstrained environment,there may exist illumination changes and pose change issues on faces.We propose a face verification method based on low-level features attribute prediction.At first,we use a variety of low-level feature sets to select discriminative features,then predict face attributes according to the obtained features,finally we train a SVM classifier to perform face verification.Experimental results on LFWA face attribute dataset show that the proposed method has obtained a higher accuracy in face attribute prediction,and also achieves 5% improvement in face verification accuracy compared with other 4 methods.Comparative experiments on cross-Age MORPH face datasets show that the proposed method has a lower equal error rate compared with other methods.3)Low-level features can only express single information and hardly contain semantic information,in order to get more discriminative attributes,we propose the method of fusing high-level abstract feature attributes and low-level feature attributes to conduct face verification.Firstly,we use self-learning abstract features extracted by deep network to train attribute classifiers,and then fuse the predictable face attributes with the low-level feature attributes researched in the third chapter by a weighted fusion method.Finally we use obtained attribute features to perform face verification by a SVM classifier.Experimental results on Celeb A and LFWA datasets show that the proposed method has further improved prediction accuracy.And compare with pure hand-crafted features,the verification accuracy has 3% improvement.Besides,experiments on the MORPH datasets show that the proposed framework obtains lower equal error rate compared to the method of the third chapter.We also perform face occlusion test on Celeb A dataset,and the results show that the proposed method is robust to occlusion.