Investigation of the performances of different deep learning models in the implementation of automatic palm print segmentation

Advances in medical imaging technologies, particularly with the use of deep learning based methods, have achieved remarkable success in specific tasks such as hand segmentation. Hand segmentation is of critical importance for identification in biometric verification systems. Accurate identification of unique biometric features such as palm vein structures, finger lengths and skin texture depends on the accuracy of hand segmentation algorithms. In this study, autoencoder based U-Net, MA-Net, FPN and PSP-Net models have been evaluated with Vision Transformer, EfficientNet-B5 and ResNet-34 backbones for solving the hand segmentation problem. The models were analyzed with metrics such as Accuracy (ACC), Intersection over Union (IoU), Dice coefficient and F1 score. The results showed that the U-Net model integrated with Vision Transformer provided the highest success in segmentation (IoU: 0.8917, Dice: 0.9892). MA-Net achieved the highest accuracy value (0.9921) when used with ResNet-34 backbone. In addition, this combination achieved the highest F1 score (0.9584). EfficientNet-B5 provided effective results with MA-Net and FPN. The results of the study revealed that the segmentation success obtained with deep learning models could increase the reliability of biometric verification systems. In particular, U-Net - Vision Transformer and MA-Net – ResNet-34 combinations showed high performance in hand segmentation. The findings emphasize that model and backbone selections are critical in hand segmentation applications.