Fine-grained Recognition Algorithm For Transformer Based On Part Features

Abstract

Fine-grained image recognition is a challenging task. Due to the small differences between the categories of fine-grained images and the large differences within the categories, traditional networks based on CNN or Transformer have their own shortcomings in feature extraction. This paper gives full consideration to the characteristics of CNN and Transformer, and proposes a fine-grained recognition algorithm combining WS-DAN (Weakly Supervised Data Augmentation Network) and ViT (Vision Transformer). Firstly, the image patch is obtained by WS-DAN to eliminate the incomplete semantic information of image patch caused by traditional ViT. Then, the image patch is encoded based on Transformer framework and global token is introduced for topological relationship constraints among components, which overcomes the locality of features extracted from traditional CNN network. Finally, the training based on the combination of cross entropy and contrast loss function further improves the recognition ability of the network. The proposed algorithm has achieved satisfactory results on the CUB-200-2011, FGVC-Aircraft and Stanford Cars datasets.