Research On Facial Attribute Recognition In Images And Surveillance

Face recognition,as part of computer vision,has arose wide interests in research,and can be applied to multiple areas.In practice,face recognition alone tends to rely much on identity information,and cannot describe the face in a direct and clear way.Facial attribute recognition can recognize one's facial features,and thus can compensate the disadvantage of face recognition when lacking identity information.In this paper,we present a novel deep learning algorithm architecture named as KT-MTL(Knowledge Transfer + Multi-Task Learning).We first exploit the correlation among facial attributes to group them,and then,according to the extent of correlation,further divide each group into the “strong” part and the “weak” part.We choose Inception-v3 as the trunk of KT-MTL,and combine it with multi-task learning structure to respectively train the “strong” attributes and “weak” attributes.Consequently,the strong net and the weak net are formed.Meanwhile,we offer to apply knowledge transfer to the algorithm.The learned information of the strong net,including parameters and loss,is used to guide the training process of the weak net.We test KT-MTL on both still images and real-time video streams.The algorithm achieves an overall accuracy of 92% on the 40 facial attributes included in CelebA dataset,and achieves the state-of-the-art accuracy of 91.89% on the “weak” part of the attributes.When performing experiments in real-time surveillance,we choose 6 measurable facial attributes for testing,and extract multiple frames of the same object in order to avoid blurring in real-time video clips.KT-MTL achieves an average accuracy of 95.76% in still images and 93.77% in real-time video.The results suggest that KT-MTL is able to achieve competitive performance in facial attribute recognition on both images and video frames.Effective and efficient,KT-MTL demonstrates its value in practical usage.