| As the country enters the digital age,massive amounts of information and data play a pivotal role in environmental protection work,and a more intelligent and digital society also puts forward higher requirements for our environmental governance.At present,one of the most effective means to solve man-made pollution from the source is to combine the mature pedestrian re-identification technology with the related direction of environmental engineering.This method can play a warning role.There are three main methods of current person re-identification technology:supervised person re-identification method,semi-supervised person re-identification method and unsupervised person re-identification method.Among them,the supervised pedestrian re-identification method is to use a deep convolutional neural network to train a large number of high-quality labeled images of environmental offenders,and obtain the best recognition model,but this process needs to consume a lot of human and material resources.The unsupervised person re-identification method does not require high-quality labeled images of environmental offenders.It uses Gan and other network structures to generate pseudo-labels for training,and its recognition accuracy is comparable to the supervised person re-identification method.Semi-supervised person re-identification methods use a small number of labeled images of environmental offenders and a large number of unlabeled images of environmental offenders to train together,and use a small amount of labeled data to achieve similar accuracy to a large number of labeled data.Therefore,by studying supervised pedestrian re-identification technology and semi-supervised pedestrian re-identification technology,and applying them to environmental engineering,this paper proposes a cross-border identification technology method for environmental offenders based on semi-supervised pedestrian re-identification algorithm.The main innovations of this paper are as follows:(1)An improved person re-identification method based on supervised learning.First,based on the position attention mechanism and channel attention mechanism in the Dual Attention Network model,an improved method of adding the position attention mechanism and the channel attention mechanism after the deep convolutional neural network is proposed,and then based on the spectrum in the Attentive but diverse Person Re Id model The Value Difference(SVDO)Orthogonality Constraint proposes an improved method to add weight constraints after each layer of convolution in the entire supervised learning network and feature value constraints after the location attention mechanism and channel attention mechanism.The experimental results show that when the deep convolutional network is Res Net50,the m AP and Rank-1 of this improved method are 88.28% and 95.96%,respectively,and the effect is better than other deep convolutional networks;this improved method can improve m AP on the basis of the baseline.and Rank-1 increased from 77.40% and91.50% to 88.28% and 95.96%,respectively.Furthermore,based on the Contextual Transformer Networks model,it is proposed to replace the convolution kernel in the Res Net network with three convolution kernels to improve the internal improvement of the deep convolutional neural network,and add weight constraints after each layer of convolution.The experimental results show that when the deep convolutional network is Res Net50,the m AP and Rank-1 of this improved method are 24.79% and44.69%,respectively;on the basis of the performance of this improved method,weight constraints are added,and the m AP and Rank-1 are changed from 24.79% and44.69% increased to 28.33% and 49.44%.(2)Semi-supervised person re-identification method based on regularization constraints.Firstly,a semi-supervised person re-identification method based on regularization constraints is proposed,which constructs a dual-model network with the same initialization parameters,namely the teacher network and the student network.During the training process,some labeled data and a large amount of unlabeled data are randomly used as the input of the dual model,and the parameters of the dual model network are constrained by regularization to keep the output results of the same labeled or unlabeled input data.Consistent.On this basis,the teacher network uses the stochastic gradient descent method to optimize the required parameters,and the student network uses the parameters optimized by the teacher network to perform the weighted moving average iteration to obtain the required parameters.Furthermore,it is proposed that the dataset in the experiment is the Market1501 dataset consisting of a large amount of unlabeled data and a small amount of labeled data.The experimental results show that the improved method has improved Rank-1 and m AP on the Market1501 dataset with some labels and a large number of unlabeled data sets compared with the original supervised learning improved method.Learning with labels is effective.Compared with unsupervised methods,semi-supervised methods and transfer learning methods,this method has advantages. |
