A novel approach to palm vein image segmentation combining multi-scale convolution and swin-transformer networks

This paper proposes a non-contact palm vein image segmentation model that integrates multiscale convolution and Swin-Transformer. Based on an enhanced U-Net architecture, the downsampling path employs a multiscale convolution module to extract hierarchical features, while the upsampling path captures global vein distribution through a sliding window attention mechanism. A feature fusion module suppresses background interference by integrating cross-layer information. Experimental results demonstrate that the model achieves 97.8% accuracy and 94.5% Dice coefficient on the PolyU and CASIA datasets, with a 3.2% improvement over U-Net. Ablation studies validate the synergistic effectiveness of the proposed modules. The model effectively enhances the robustness of palm vein recognition in complex illumination and noisy environments.