Research On Face Attribute Recognition Algorithms Based On Deep Neural Networks

Recently,face attribute recognition has become one of the research hotspots in the field of computer vision and pattern recognition.Given a facial image,the task of face attribute recognition aims to classify multiple facial attributes,such as gender,smile and charm.Face attribute recognition has a wide range of applications including image retrieval,face recognition,micro-expression recognition,image generation and recommendation systems.However,in reality applications,variations in pose and illumination as well as occlusion still make face attribute recognition a big challenge.Compared with traditional algorithms based on manually-designed features and support vector machine,deep learning-based algorithms not only avoid tedious manual feature extraction,but also jointly train multiple attributes in a unified network framework to achieve end-to-end training and prediction.The features of deep learning are more robust and deep learning-based algorithms are more time saving.So,the research on deep learning-based face attribute algorithms has important theoretical value and practical significance,and has become the mainstream research direction in the field of face attribute recognition.The main works in this thesis are summarized as follows:(1)We propose a novel deep multi-task multi-label convolutional neural network,termed DMM-CNN,for effective FAC.Specifically,DMM-CNN jointly optimizes two closely-related tasks(i.e.,facial landmark detection and FAC)to improve the performance of FAC by taking advantage of multi-task learning.To deal with the diverse learning complexities of facial attributes,we divide the attributes into two groups:objective attributes and subjective attributes.Two different network architectures are respectively designed to extract features for two groups of attributes,and a novel dynamic weighting scheme is proposed to automatically assign the loss weight to each facial attribute during training.Furthermore,an adaptive thresholding strategy is developed to effectively alleviate the problem of class imbalance for multi-label learning.Experimental results on the challenging CelebA and LFWA datasets show the superiority of the proposed DMM-CNN method compared with several state-of-the-art FAC methods.(2)We propose a novel facial attribute recognition method based on the generative adversarial network and self-supervise learning,termed GAN-SSL.First,we train a face attribute manipulation network to enhance the training data of facial attributes.Then,we train a self-supervise learning network,which can recognize the rotation different types.The self-supervise learning network can be trained without using attribute labels.Finally,we fine-tune the self-supervise learning network to perform face attribute recognition task with a few fake and real face images.Experimental results show that,when we use 1/10 data as the training data on CelebA,LFWA and UMDUED datasets,our method can improve the mean accuracy by 2.42%,3.17%and 5.77%,respectively.Our method can effectively alleviate the lack of labeled data in face attribute recognition and remarkably improve the performance especially on small datasets.