Deep Face Images Hashing With Generative Adversarial Networks

Face Image retrieval is one of the key research directions in computer vision field.Image hashing technology has become a research hotspot because of its high storage efficiency and fast retrieval speed.Thanks to the rapid development of deep neural networks in recent years,deep hashing has achieved good performance in the field of image retrieval.But for large-scale face retrieval,the performance needs to be further improved.In this paper,we propose Deep Face Hashing with GANs(DFH-GANs),a novel deep hashing method for face image retrieval.Considering the characteristics of deep learning and hashing methods,this method makes some improvements on several aspects such as loss function,quantization method,training data and network structure.This paper mainly including the following research work:(1)In order to solve the problem that the existing loss functions in deep hashing algorithms for face images does not discriminative enough,this paper proposes a face hash algorithm based on metric learning.This hash algorithm introduces the idea of metric learning in the hash field.At the same time,it optimizes the characteristics of the hash code and controls the sign of the hash code to help the network learn more accurate hash code.(2)In order to solve the problem that the network convergence is slow and the generated hash code is not robust because of the existing quantization methods,this paper proposes a face hash algorithm based on the loss of quadratic quantization.This paper analyzes the gradient of the quantization method during back-propagation and designs a "quadratic quantization loss".This loss makes the network converge faster and better,improve the robustness of the network,and reduce the quantization error.(3)In order to solve the problem of insufficient training samples of single human face images and the loss of information in the information compression step of the hash method,this paper uses GANs for face attribute data enhancement and proposes a multi-layer supervised DFH structure and a multi-stage training method.This paper draws on the idea of knowledge distillation and proposes a multi-layer supervised DFH structure.Through multi-stage training,it helps the network to learn better features and thus learn more discriminating hash codes.Experiments on multiple datasets show that DFH-GANs can generate high-quality binary hash codes,and has achieved good face image retrieval effects.