Human Dorsal Hand Vein Segmentation Method Based on GR-UNet Model

To solve the issue of inaccurate segmentation accuracy of human dorsal hand veins (HDHV), we propose a segmentation method based on the global residual U-Net (GRUnet) model. Initially, a visual acquisition device for dorsal hand vein imaging was designed utilizing near-infrared technology, resulting in the creation of a dataset comprising 864 images of HDHV. Subsequently, a Bottleneck from the deep residual network-50 (ResNet50) was integrated into the U-Net model to enhance its depth and alleviate the problem of vanishing gradients. Furthermore, a global attention mechanism (GAM) was introduced at the junction to improve the acquisition of global feature information. Additionally, a weighted loss function that combines cross-entropy loss and Dice loss was employed to address the imbalance between positive and negative samples. The experimental results indicate that the GR-Unet model achieved accuracies of 78.82%, 88.03%, 93.92%, and 97.5% in terms of intersection over union, mean intersection over union, mean pixel accuracy, and overall accuracy, respectively.