Abstract: To solve the classification difficulties of small inter-class differences and large intra-class differences in fine-grained classification, an improved end-to-end fine-grained classification model (TBformer) is proposed based on Swin Transformer. In view of the interference of complex background on network recognition, the dynamic location module (DLModule) combining ECA, Resnet50 and SCDA is used to capture key objects, and a three-branch feature extraction module based on DLModule is designed to improve the ability of target discriminant feature extraction. In order to fully tap the rich fine-grained information contained in the three-branch features, a feature fusion method based on ECA is proposed to enhance the comprehensiveness and accuracy of the features, and improve the robustness of the network for fine-grained classification. The experimental results show that compared with the basic method, the accuracy of TBformer is improved by 3.19% in CUB-200-2011, 3.47% in Stanford Dogs and 1.09% in NABirds.

Key words: fine grained recognition, feature fusion, attention mechanism, multiple branches