Research On Large-Scale Face Image Retrieval Based On Deep Hashing

While enjoying the convenience brought by the rapid development of Internet technology,people also generate and accumulate a large number of images on the Internet.How to effectively organize and mine useful information at a deep level in the image is a problem that we have encountered.Content-based image retrieval technology has developed rapidly and has been widely used in social,security,and people's livelihoods.However,with the exponential growth of image size,the technology faces great challenges in terms of retrieval precision and retrieval delay.The hashing method is favored for its low memory occupation and high retrieval efficiency.At the same time,in the sake of the outstanding achievements of deep learning technology in the field of computer vision,the hash method based on deep learning has gradually entered the field of view of researchers.The existing depth hashing method usually separates the feature extraction and quantization process,focuses only on the similarity relationship between pair-wise images and ignores the semantic information of the picture itself,ignoring the redundant information between the quantized feature bits.This thesis aims at improving the existing hashing method,and proposing a deep hashing method that combines the similarity between the picture and the semantic information of the picture itself,and designs and implements a GPU-based multi-level parallel retrieval scheme.While improving the retrieval accuracy,the retrieval speed on large-scale data sets is accelerated.The specific work of this thesis is as follows:(1)Proposed a deep hashing method.Based on the existing pair-wise hashing method,a multi-task deep learning mechanism is used to combine the classification loss function with the contrast loss function.During the quantization process,the similarity between the image pairs is preserved,and as much as possible.To retain the semantic information of the picture itself,classification tasks and quantitative tasks guide each other's learning.Finally,an end-to-end feature extraction and quantification network was obtained.A large number of experiments on CIFAR-10 and ms-celeb datasets have proved the effectiveness of the method.(2)The local connection module is used instead of the full-connection layer of the quantized network.Each quantization bit is related to only part of the input,for reducing the redundant information between the features.Finally,2-3 percent m AP were improved on both data sets.(3)Learning from Res Net,we design and implement a deeper network.Deep networks can usually get good feature expressions,while high retrieval accuracy usually depends on good feature expressions.Experimental results show the effectiveness of the network.(4)On the basis of Hamming sorting,a multi-level parallel retrieval method based on GPU is designed and implemented.This method integrates multiple retrieval tasks into one task and uses GPU natural parallelism to complete the retrieval.Finally,the effect of delaying the single-page retrieval by 0.8ms in a million-scale image library was achieved.