Adversarial Contrastive Learning Based on Image Generation for Palm Vein Recognition

Vein recognition has gained significant attention in research due to its emphasis on security and privacy. Deep learning methods have recently achieved remarkable success in the field of vein recognition. However, deep learning methods are unable to achieve optimal results on vein datasets due to their limited size. To overcome this limitation, we propose Adversarial Contrastive Learning (AdCL), a method that combines UNet with contrastive learning. AdCL utilizes UNet to generate challenging samples through adversarial means, and then inputs these samples into the contrastive learning model to learn more robust feature representations. Throughout this process, the weights of both UNet and the encoder are updated alternately. Initially, the UNet takes the augmented images obtained through classic data augmentation as input. Subsequently, the generated images are employed to train the contrastive learning model. Following training, the trained encoder is extracted and combined with the classifier to perform vein image classification. Experimental results on two datasets of palm vein images demonstrate that our proposed AdCL method enhances classifier performance and surpasses existing contrastive learning approaches.