Vehicle Re-identification Approach Combining Multiple Attention Mechanisms and Style Transfer

To address the impact of inter-domain style differences on cross-domain performance degradation in vehicle re-identification applications, a vehicle re-identification method combining multiple attention mechanisms and style transfer is proposed. Firstly, IBN-Net is introduced on the basis of ResNet50 for improving the generalization ability of the network, secondly, EFDMix method is used to generate more diverse feature enhancements to improve the domain adaptation ability of the model, while multiple attention mechanisms are combined to extract more discriminative vehicle features and improve the feature representation ability of the model on different datasets, and finally, cross-entropy loss with label smoothing and supervised contrastive loss for vehicle sample classification and optimising the feature distance between categories. Single-domain vehicle re-identification experiments were conducted on two datasets, VeRi-776 and VehicleID, and the results showed that our proposed method outperformed existing methods in most performance metrics. Further cross-domain comparison experiments also demonstrate the advancedness of the proposed method in terms of cross-domain.