Fine-Grained Visual Categorization with Window Restrain for Intelligent Products

Fine grained visual classification (FGVC) refers to the classification of subcategories in a dataset that belong to the same basic category. The difficulty lies in the less inter class sample differences and the more intra class sample differences, which are more challenging compared to traditional classification tasks. Recently, the use of Vision Transformer (ViT) methods has achieved significant results in FGVC scenes. However, mainstream research methods are based on the enhancement idea of local features, ignoring the interference of irrelevant background features on local discriminative regions. Therefore, in order to address existing deficiencies of Vision transformers being susceptible to complex background interference and insufficient ability to extract local features in fine-grained classification scenes (FGVC), this paper proposes a window restrain module (WRES) and combines it with the backbone network of standard Vision Transformer, for the task of fine grained classification of intelligent products, effectively enhancing the Vision Transformer model's ability to extract features from local discriminative regions and achieving end-to-end training. The results of experiments show that the accuracy of the Vision Transformer network incorporating WRES modules in CUB-200-2011 reaches 92.7%, which is 2.4% higher than the standard ViT method. It also outperforms the current fine-grained visual classification methods based on CNN and Vision Transformer on other FGVC benchmark datasets and intelligent products FGVC tasks.