| With the popularity of surveillance cameras and the rapid development of graphic detection technology,"shoe print and surveillance" technology has become one of the important means for public security organs to solve cases.This technical warfare method uses the suspected shoe prints on the scene to query the corresponding shoe type,and then finds the same shoe type in the surveillance video,so as to lock the suspect,which plays an important role in criminal investigation.However,at present,the search for suspected shoe types in the surveillance video is completed manually,which consumes extremely human and material resources,and is easy to be missed due to subjective factors.Therefore,there is an urgent need for an automatic shoe shape recognition method to further enhance the practicability of this technical warfare method.Based on this,the research on automatic shoe shape recognition method based on convolutional neural network is carried out.The main work is as follows:Firstly,aiming at the problem of lack of data set,a multi background shoe data set under surveillance video is established.A surveillance camera is set up to simulate the actual combat of public security criminal investigation and collect surveillance video data.After a series of operations such as framing,interception and preprocessing,a shoe shape data set containing35300 images of 300 types of shoes is established,in which the training set contains 30000 images of 150 types of shoes and the test data set contains 300 images of 150 types of shoes.In addition,5000 confused sample shoes and the test data set are selected to form a database to test the performance of the algorithm.Secondly,aiming at the problem that the network could not extract shoe type features effectively,a shoe type recognition algorithm based on attention mechanism and feature layer selection was proposed.Based on the characteristics of the built data set,an attention mechanism model is designed to enhance the ability of network feature extraction.The output of layer conv4＿x,layer conv5＿x and layer FC of Res Net50 were extracted respectively as shoe features to compare the recognition accuracy,and it was found that the output of convolutional layer as shoe type features had better recognition effect.The test results on shoe shape data set show that the accuracy of Rank-1 algorithm is improved from 61.09% to 74.32%,and the accuracy of m AP is improved from 46.34% to 56.97%.Thirdly,aiming at the problem that the algorithm has poor recognition effect on the shoe shape without obvious marks and more details are lost,a shoe shape recognition algorithm based on adaptive sensing field and multi-branch feature fusion is proposed.An adaptive receptive field module is designed for network adaptive selection of appropriate receptive field features to further enhance feature extraction capability.A three-branch feature fusion model is designed to make the network make full use of feature information and improve the accuracy of network recognition.The test results on shoe shape data set show that the accuracy of Rank-1 is improved from 74.32% to 78.21% and the accuracy of m AP is improved from 56.97% to 61.70%on the basis of the improved scheme in Chapter 3.Fourthly,aiming at the shortcomings of the original Res Net50 network structure,the bottleneck structure and the down-sampling module are improved to strengthen the effective transmission and flow of feature information in the network,and a Strong baseline is proposed combining with the data enhancement method.Finally,the Strong Baseline is integrated with the improvement scheme in Chapter 4 and the shoe shape automatic recognition model（STRNet）is proposed.The test results on the shoe shape data set show that the Rank-1 and m AP accuracy of STRNet reach 86.38% and 66.58%,which are 25.29% and 20.24% higher than original Res Net50 network,respectively. |
PDF Full Text Request