Research On Person Re-identification Method Based On Deep Learning

Person re-identification(ReID)is the use of image processing technology to retrieve images of designated person in the database given the need to query person.The current data set used to train the person re-identification model comes from a limited fixed collection device,and the diversity of sample styles is lacking.At the same time,because the crossdomain person re-identification technology has a large sample difference between the source domain and the target domain,the person re-identification model trained in one domain is directly tested on the other domain,and its performance obviously fails to meet expectations.In view of the difficulties in the above situation,this article adopts the method of ReID based on deep learning to implement the study.Mainly use relevant algorithms and theoretical techniques in deep learning to carry out research on person re-identification and cross-domain person re-identification.First,a person re-identification method for crossdomain image style transfer is proposed.This method uses an improved cyclic generation of confrontation network framework,which can transfer the style and style of image data captured by different cameras from different data sets,and solve the existing problem.In the cross-domain person re-identification technology,the source domain and target domain samples are quite different,and the test effect of cross-domain person re-identification is improved.Secondly,a person re-identification method based on single-domain image style transfer based on positive and negative sample training is proposed.It also uses an improved cycle generation confrontation network framework,which can make the image data captured by different cameras in the same data set perform Style transfer can improve the diversity of single-domain image sample styles at a lower cost.This method uses the data after the style transfer as a negative sample,and the data before the style transfer as a positive sample.The positive and negative samples are sent to the model training at the same time,and while controlling the input ratio of the positive and negative samples,different loss functions are used to obtain Optimize the loss of positive and negative samples to improve the generalization ability of the model.The contributions of this article in academic and practical applications are as follows:1.In order to smooth the lack of diversity of sample styles in ReID technology within a single domain,and the large difference within samples between the source domain and the target domain.In this paper,through the improvement of the cyclic generation confrontation network,unpaired image data can be directly sent to the model for training.This framework enables the source domain image to combine the image style of another domain,and its similarity is closer to the similarity of the person image data in the target domain.The sample diversity is enhanced than before conversion,and the sample difference between domains is reduced.And make the image data captured by different cameras in a single domain undergo style transfer,which can improve the diversity of sample styles in a single domain at a lower cost.2.In order to smooth the issue of cross-domain ReID technology,the source domain and target domain samples are quite different,resulting in a significant drop in performance when a model trained in one domain is directly tested on another domain.This paper adopts a cross-domain image style transfer person re-identification method to increase the sample diversity,which reduces the difference between the source domain and target domain samples.Specifically,through the use of an improved cyclic generation of confrontation network framework,the similarity of the source domain image combined with the image style of another domain is closer to the similarity of the target domain data,and the sample diversity is enhanced than before the conversion.Which reduces the sample difference between domains and makes the model have better generalization performance.The test results show that by performing style transfer on the Market-1501,Duke MTMC-re ID and MSMT17 data sets,and then extracting global features through the residual network,the accuracy of cross-domain recognition is significantly improved.At the same time,the model can achieve a better re-recognition effect without combining the style and style of the target domain data set or in the case of a small target data set,which is better than other crossdomain person re-identification methods with better current tests.3.In order to smooth the issue of lack of diversity in ReID sample styles within a single domain,this paper uses an improved cyclic generation confrontation network framework to transfer the style and style of the image data captured by different cameras in a single domain,which can be achieved at a lower cost.The diversity of sample styles has been improved.A new training mechanism of positive and negative samples fusion is designed.First,the samples after the style transfer are used as negative samples,and the samples before the style transfer are used as positive samples.The positive and negative samples are simultaneously sent to the model according to different proportions for training.Further,in order to prevent overfitting and consider the loss of wrong label positions,label smoothing regularization is adopted.At the same time,in order to pay more attention to difficult and error-prone samples,and to optimize the loss of negative samples,a focal loss function is adopted.The model trained by reducing costs and combining the training mechanism of positive and negative samples has stronger generalization ability,and the effect of person re-identification is better.Significantly improved 1.51% and 2.07% on the Market-1501 data set and Duke MTMCre ID data set,respectively.In summary,this article uses deep learning technology to carry out research experiments and analysis on person re-identification tasks from the perspective of sample differences between domains and sample diversity within a single domain.These technologies can be applied in related fields such as robot vision.