Convolutionally Enhanced Feature Fusion Visual Transformer for Fine-Grained Visual Classification

Fine-grained image classification is a popular research topic in computer vision and pattern recognition, where the goal is to recognize and classify subclasses of objects in images at the fine-grained level. In recent years, Transformer's self-attention mechanism has been increasingly introduced into fine-grained image classification tasks due to its ability to naturally focus on the most discriminative regions of the object. In this paper, a new Convolutionally Enhanced Feature Fusion Visual Transformer method is proposed based on the Feature Fusion Visual Transformer by introducing convolutional operations. Firstly, for the original input image, patches are not directly labeled, but extracted from the generated low-level features; Secondly, the computational complexity at the multi-head attention layer is reduced through spatial-reduction attention, which also reduces memory consumption; Finally, the inverted residual feed-forward network is applied to each encoder to improve the network's expression ability. Comparative experiments on four datasets show that the method improves the accuracy of fine-grained image feature extraction and reduces the computation and memory consumption by improving the self-attention layer to improve the efficiency and performance of the model.