Invariant Cross-domain Person Re-identification Based On Automatic Data Enhancemen

Pedestrian re-recognition technology aims to retrieve the target pedestrian from a database of pedestrian images taken by multiple cameras.Pedestrian re-recognition models based on deep learning can achieve good performance when they are supported by a large amount of data.However,most of the work in the field of pedestrian re-recognition mainly focuses on designing better network architecture models,and few people have explored the design of better data enhancement methods.In practical application scenarios,pedestrian re-recognition also faces cross-domain challenges.That is,when a pedestrian re-recognition model with good performance in one dataset is applied to another dataset,its performance will be greatly affected because the source domain data and target domain data come from different scenarios.At present,the problem of image difference between source domain and target domain is mainly solved by reducing the difference between the two domains.However,these methods do not further consider that the model can be adapted to different pedestrian re-recognition data sets by improving its own structure.In addition,these methods tend to ignore the intra-domain variation of the target domain,but the intra-domain difference within the target domain often has an important impact on the model performance.Therefore,this paper studies the problems in cross-domain pedestrian re-recognition,such as difficult data acquisition,large difference in image appearance between source domain and target domain,and intra-domain variation of target domain.The main research contents are as follows:(1)In view of the difficulty in obtaining cross-domain pedestrian re-recognition data sets,automatic data enhancement is used to enhance the data sets of source and target domains.Automatic data enhancement framework mainly consists of two parts:search algorithm and search space.The search space mainly contains relevant data enhancement operations,and the search algorithm is the optimal strategy used to find the data enhancement operations.In the search process,the hyperparameters are optimized based on Bayesian optimization,which is a method to find the minimum or maximum value of the objective function with the guidance of Bayesian theorem.Experimental results show that automatic data enhancement can find the best enhancement strategy for the data set,and has good migration,and can effectively increase the image samples in source domain and target domain.(2)For cross-domain pedestrian re-recognition,the data images in source domain and target domain differ greatly in style,style and other appearance,and the images in target domain change.Instance normalization and batch normalization are added to the network model to improve the discrimination of the network to the image,and in-domain invariance is added to the target domain to solve the in-domain variation.Experimental results show that adding appearance invariance to the model can improve the adaptability of the model to image appearance changes and improve the generalization ability of the model.