Internet Picture Face Retrieval System Based On Deep Learning

Recent years,face retrieval technology has penetrated into the national security,finance,civil affairs,monitoring and other fields.On the one hand,with the explosive growth of Internet face images,how to retrieve the relevant faces from the massive human face datasets quickly has become a hard problem,and the approximate neighbor search method represented by hash has been widely concerned and studied.On the other hand,convolutional neural networks have achieved amazing performances in computer vision areas.Thus,the deep hash algorithm gets favor in researchers' eyes.The existing deep hash algorithms often neglect the change of the data distribution brought by the binarization process,and usually can not fully utilize the strong feature extraction ability to promote the hash function learning.Also,the convolutional neural network needs the expensive calculation cost and huge parameter storage space.Based on these reasons,this paper proposes a deep hash algorithm based on classification and quantization error,and further compresses the deep network,which can accelerate the calculation efficiency and reduce the parameter storage space while guaranteeing the accuracy of networ.In the end,we design and implement an Internet image face retrieval system.The concrete work of this paper is as follows:1)We propose a novel face retrieval algorithm based on deep hash.We take advantage of the deep neural network to extract effective features,and combine deep network and hash coding,using both prediction and quantization error to directly guide the training of network,which helps to keep the clustering distribution of original data and reduce the quantization error,so as to obtain the hash coding with accuracy and effectiveness.2)We take use of deep compression algorithm to compress the proposed network model.As the network model continues to expand in terms of depth and width,the computational cost and parametric storage space of the model increases exponentially.In this paper,the compression of parameter storage is realized from three aspects:network pruning,weighting sharing and Huffman coding.The network compression ratio of 35x-49x is achieved by implementing the parameter compression on the LeNet-5 network,AlexNet network,VGG-16 network and the DHCQ network proposed in this paper under the basic premise of without network accuracy loss.3)We design and implement a face retrieval system.The system can let users easily browse the face retrieval results,and will give a specific evaluation named MAP in the current retrieval situation,to facilitate users to evaluate the final results of face retrieval from both objective and subjective aspects.