An Improved Swin Transformer-based Model for Fine-grained Image Classification

Fine-grained classification is a challenging task. Different from traditional classification, there frequently arises the issue of significant intra-class variation and minor inter-class variation in fine-grained classification, which poses great difficulties for classification. Recently, the attention-based Swin Transformer has surpassed traditional CNN models in various visual tasks. It can effectively capture global features and achieve better classification performance. In contrast, convolutional neural networks are adept at extracting local features. However, fine-grained classification tasks rely more on local features. Therefore, this paper attempts to further improve the features extracted by the Swin Transformer and subsequently fuse them with the low-level features extracted by CNN, thereby forming a feature representation more suitable for fine-grained classification. Eventually, competitive results have been obtained on common fine-grained classification datasets.