Research On Unsupervised Cross-domain Person Re-identification Combined With Attention Mechanism

Person re-identification technology is widely used in smart security and social safety issues such as searching for lost children.Existing person re-recognition techniques mainly include supervised person re-recognition and unsupervised person re-recognition.Supervised person re-recognition mainly learns the information of person images with labels and achieves good results.However,labeling the dataset is costly,so unsupervised person re-identification of scenes is increasingly important in practical applications where persons appear in different domains.Compared with supervised methods,unsupervised person re-identification learns information from source domain data with labels and target domain data without labels,and usually,the two datasets are different.The model trained on the source domain data set is affected by environmental factors and lacks scalability when applied directly to the target domain data set,and the model has significant performance degradation.In this thesis,the unsupervised cross-domain person re-identification method is investigated by the following main research work.(1)To address the problems of lack of feature discrimination and clustering generating pseudo-label noise.First,when extracting person features,coordinate attention block and triple attention block are concatenated and added to the backbone network Res Net-50 for feature aggregation to alleviate the loss of location information caused by global pooling,capture the cross dimension to calculate attention weights,and my fine-grained information.Secondly,DBSCAN clustering is improved to design reliability metrics to determine the reliability of clustered instances and realize the clustering process from coarse to fine.Finally,the clustering centers are dynamically stored in the memory module,and the class centers are dynamically constructed for invariant learning.The experimental results show that the new method designed in the thesis improves the accuracy of m AP by 2.4% on the target domain Market-1501 dataset and 6.5% on the target domain Duke MTMC-Re ID dataset,respectively,compared with other typical methods.(2)For the two problems of ignoring local features and feature variations caused by domain gaps.First,invariant and specific features from different domains are modeled to alleviate the domain gap problem by fully considering the diversity and complementarity of features from different domains.Second,global features and two local features are clustered independently in each training cycle to fine-tune the whole network using refined pseudo-labels for each branch.Finally,the meta-learning optimization module helps the model to learn camera-invariant representations,and the model is optimized using the proposed meta-learning strategy to encourage the model to learn camera-invariant features.The experimental results show that the new method designed in the thesis improves the accuracy of m AP by 2.0% on the target domain Market-1501 dataset and m AP by 2.5% on the target domain Duke MTMC-Re ID dataset,respectively,compared with other typical methods.