Face Retrieval Method Based On Deep Hash

With the growing popularity of social networking on intelligent mobile services,the number of images containing faces has witnessed an explosive increase.Consequently,face image retrieval,which aims to identity images containing the person in the given face images,has become an attractive research area.The main challenges of face image retrieval are large intra-class variations and the large cost of computing time and storage.In order to solve the first problem,we adopt the facial feature extraction method based on deep learning.For the second problem,we use the hashing method to solve it.Therefore,this paper uses deep hashing method with a combination of deep learning and hashing method to solve the task of face retrieval.However,these methods mainly adopt the pairwise or the triplet ranking as their loss function.It is cost a lot of time and space to obtain and store these pairs or triplets.Moreover,the optimization complexity of models would greatly increase and the learned hash codes cannot guarantee to be discriminative.In view of this,this paper use simple classification loss function instead of pairwise or the triplet ranking loss,and incorporate the feature extraction and hash learning into a unified framework to construct a end-to-end system.After that,a lot of improvements have been made to improve the retrieval accuracy.In addition,a large-scale face retrieval system was designed and implemented in this paper.The main research work of this paper is as follows:(1)This paper proposes a simple yet effective method for scalable face image retrieval based on deep discriminative hashing(DDH)algorithm.The proposed framework leverages feature extraction,hash learning and class prediction in one unified network to construct a end-to-end system.And the divide-and-encode module is introduced to reduce the redundancy among hash codes and the network parameters simultaneously.The proposed method can learn discriminative and compact hash codes.Finally,extensive experiments on two widely-used facial datasets,YouTube Faces and FaceScrub,demonstrate that the proposed method achieves superior performance compared with other hashing methods.(2)In order to solve the existing problems in the DDH,we propose a discriminative deep quantization hashing(DDQH)algorithm.This method has made many improvements in the aspects of facial feature extraction,quantitative structure and multiscale feature fusion,which greatly improves the retrieval accuracy of the network.Moreover,those experiments on YouTube Faces and FaceScrub proved that our DDQH method has achieved good results on both simple and complex facial datasets.(3)In this paper,a large-scale face retrieval system is designed and implemented.The user can select different face datasets in this system,and then select the facial image to be retrieved by clicking or uploading,after that,the system will quickly return image retrieval results and the corresponding retrieval accuracy,so that users can evaluate the feedback of the system effectively.