Vision transformer with pyramid feature extraction and efficient channel attention for classifying images

Existing image recognition networks face challenges in achieving satisfactory accuracy and robustness on complex image datasets. Vision Transformer (ViT) has demonstrated unique potential in addressing these challenges. However, the fixed-size patch embedding strategy used in standard ViT limits their capacity for multi-scale feature extraction. To overcome this limitation, this paper proposes a novel high-performance image recognition architecture by integrating a pyramid-based patch embedding strategy, an efficient channel attention (ECA) module, and an optimized residual connection structure. The pyramid embedding method enables the simultaneous capture of local and global features across different scales, while the ECA module enhances sensitivity to important channel features. Furthermore, the optimized residual connection structure improves feature propagation efficiency within the Transformer encoder. Experimental results on the CIFAR-10 dataset show that the proposed method outperforms the ResNet18 baseline, achieving higher classification accuracy and computational efficiency. Ablation studies validate the effectiveness of each component and offer new directions for the design of ViT.