Optimization Of Convolutional Neural Networks And Its Application In Pedestrian Re-identification

With the continuous development of software and hardware technology and the arrival of the era of big data,deep learning has become a research hotspot.Excellent results have been achieved in the fields of image recognition,speech recognition and natural language processing.Convolutional neural network is a commonly used deep learning algorithm and is one of the well-developed deep learning methods.Compared with the traditional method,this method can independently learn and classify efficient features.However,the convolutional neural network is too deep,and there are too many network parameters,which may cause the network training speed to slow down and it is difficult to converge.In order to solve the problems that often occur in the network,a network with higher recognition rate and fewer parameters is designed.By combining the classical convolutional neural network structure design structure and continuously optimizing the parameters,a good performance training model is obtained,which achieves the purpose of improving recognition rate and reducing parameters.Combined with the basic theory of convolutional neural network,it is applied to pedestrian re-identification,and the network model is applied to the real life field to verify its validity and significance.The main research contents and innovations of this article are as follows:(1)The basic concepts of convolutional neural networks are introduced,including working principle and basic network structure.The Caffe framework,data preprocessing methods and data sets used in deep learning are briefly described.(2)For the traditional convolutional neural network,the problem that the extraction feature information is insufficient,the classification accuracy is poor,and it is easy to over-fit.A convolutional neural network based on multi-branch aggregation is proposed.Based on the traditional convolutional nerve,this network increases the width and depth of the network without increasing the parameters,optimizes and improves the network,further improves the feature expression ability of the network,enriches the diversity of extracted features,improves the accuracy of image classification,and prevents over-provisioning.Hehe.This network,traditional network and other networks were compared and analyzed through a series of comparative experiments in two public data sets CIFAR-10 and CIFAR-100,and the effectiveness of the network was proved.(3)Although the classification accuracy is good for the traditional deep convolutional neural network,the parameter quantity is huge.which is difficult to deploy in memory-constrained devices,a multi-scale parallel fusion lightweight convolutional neural network architecture PL-Net is proposed.Firstly,the upper output feature map is sent to two different scales of the depth separable convolution layer,and then the parallel output is cross-fused with the feature information,and the residual learning is designed to design a parallel lightweight module PL-Module;To better extract the feature information,the scale-dimensional reduction convolutional module(SR-Module)is proposed to replace the traditional pooling layer.Finally,the above two modules are stacked on each other to construct a lightweight network.In the experimental phase,training and testing were performed on the CIFAR-10,Caltech-256 and Food-101 data sets.The results show that compared with the traditional CNN,MobileNet-V2 and Squeezenet networks of the same scale,PL-Net improves the classification accuracy of the network while reducing the amount of network parameters,and is suitable for deployment on memory-constrained devices.(4)Pedestrian re-identification is a research direction in the field of intelligent vision.However,due to the uncontrollability of the shooting environment,the difficulty of research is caused.Therefore,the convolutional neural network is applied to the field of pedestrian re-identification,using a lightweight network.Improve the accuracy of pedestrian re-identification,and evaluate the performance of the proposed method with the other four network models on the test set of two public pedestrian re-identification data sets of Market1501 and Duke.Finally,a visualization tool for pedestrian re-identification was written to clearly show the effects of the model.