Face Detection And Large-scale Object Retrieval For Crowd Supervision

With the rapid development of modern information technology,the intelligent video monitoring technology has become an important part of the modern information construction for the public security system.Through HD cameras deployed in shopping malls,railway stations,bus stations and other important places,the public security system can conduct real-time monitoring on the information of people in public places,to supervise the security dynamics in key areas,to detect and track high-risk personnel and to increase the work efficiency of urban security administration.Yet as the scale of the video monitoring system grows rapidly,the video management system for public security based on attended operation can no longer meet the demands of massive video image information analyses.Therefore,fast information retrieval in large-scale image resources has become one of the most important issues to be solved in the video monitoring system for public security.In the paper,to meet the demands of public security information system on monitoring image retrieval,a large-scale image information retrieval technology on account of deep learning and Hashing coding has been put forward based on existing image retrieval technologies.Main contents of the paper are listed as follows:Firstly,the facial feature extraction technology based on deep learning has been studied to solve the low image feature extraction precision of the large-scale image retrieval technology.A large-scale face recognition image retrieval method based on Convolutional Neural Network(CNN)has been put forward,which adopts CNN to acquire high-grade image features so as to improve the precision of the whole retrieval architecture.Secondly,based on the large-scale image retrieval method of CNN,index construction of Hashing coding has been carried out to optimize the rate of retrieval system by combing the image features extracted from every level of CNN with Hashing.Experimental simulation findings show that the retrieval precision of several data sets including MNIST,CIFAR-10 and FERET has reached over 90%and the speed of image retrieval has been improved.To accelerate the image processing speed,parallel acceleration has been conducted on CNN.Experimental analyses show that it takes 32.16 minutes for CNN with a parallel acceleration to process 50,000 RGB images in a size of 32*32.Instead,it takes 42.13 hours for CNN without a parallel acceleration.In comparison with existing methods,the speed has been increased by about 79 times.Thirdly,a community supervision system in pilot regions of Fuzhou has been constructed based on the large-scale image retrieval technology proposed in the paper,having realized multiple functions including auxiliary entrance guard,retrieval and reverse retrieval.According to the analyses of practical operation,the retrieval precision of the system can reach more than 90%.It takes about 0.063 seconds for retrieving one image averagely,which has basically met current demands for crowd supervision.