Research On The Methods Of Deep Neural Network Based Facial Attribute Classification

Facial attribute classification is an important research area in computer vision and pattern recognition,which has received significant attention.The main task of facial attribute classification is to predict the attributes of a given facial image,including smile,gender and attraction,etc.Facial attribute classification shows widespread and practical applications in face verification,face recognition and image retrieval,etc.However,it remains a great challenge,due to the large facial appearance vari-ations caused by pose,illumination and occlusion,etc.On the other hand,deep learning has achieved outstanding performance in a variety of computer vision tasks,including facial attribute classification.Deep neural network is a very hot topic in deep learning in recent years.Therefore,researches on deep neural network based facial attribute classification are meaningful and challenging.The detailed work in this paper is summarized as follows:(1)We propose a multi-task learning of cascaded CNN for facial attribute classification method.Conventional facial attribute classification methods usually firstly pre-process the input images(i.e.,perform face detection and facial landmark localization)and then predict facial attributes.These methods ignore the inherent dependencies among these tasks(i.e.,face detection,facial landmark localization and facial attribute classification).Moreover,some methods using convolutional neural network are trained based on the fixed loss weights without considering the differences between facial attributes.In order to address the above problems,we propose a novel multi-task learning of cascaded convolutional neural network for facial attribute classification.We achieve 91%of the mean average precision on the CelebA dataset and achieve 84%of the mean average precision on the LFWA dataset.(2)We propose a multi-label learning based deep transfer neural network for facial attribute classification method.Conventional facial attribute clas-sification methods based on deep neural network depend on a massive amount of labelled data.However,in real-world applications,labelled data are only provided for some commonly used facial attributes(such as age,gender);whereas,unlabelled data are available for other attributes(such as attraction,hairline).Moreover,con-ventional facial attribute classification methods ignore the correlation between the facial attributes.To address the above problems,we propose a novel deep transfer neural network method based on multi-label learning for facial attribute classifica-tion in this paper.To take advantage of the correlation between the facial attributes,we propose an effective loss weight scheme to compute the loss weights of facial at-tributes based on attribute grouping for facial attribute classification.We achieve 92%of the mean average precision on the CelebA dataset and achieve 84%of the mean average precision on the LFWA dataset.