Cross-Modality Domain Adaptation for hand-vein recognition

Palm-vein recognition has attracted increasing attention over the last years. Although deep learning-based approaches, such as Convolutional Neural Networks (CNN), have been shown to be effective for feature representation, thereby achieving good performance in vein verification tasks, they typically are trained on large labeled datasets. In general, labeling vein images is expensive and time cost, and typical hand-tuned approaches for data augmentation can not collect the complex variations in such images. To address this problem, a novel unsupervised domain adaptation approach, named CycleGAN-based domain adaptation (CGAN-DA), is proposed to automatically extract discriminant from the palm-vein network, without the need of any image annotation. Our proposed CGAN-DA allows a learning scheme that ensures a synergistic fusion of adaptations image-wise and feature-wise. Concretely, we transform the image appearance across two domains (palm-vein image domain and retinal image domain), in order to enhance the domain-invariance of the extracted features for the palm-vein segmentation task. Without using any annotation from the target domain (palm-vein images), our model learning is guided by several adversarial losses, a cycle consistence loss and a segmentation loss. Our experimental on the public CASIA palm-vein dataset show that our approach is capable of achieving state-of-the art verification accuracy.