TripletGAN VeinNet: Palm Vein Recognition Based on Generative Adversarial Network and Triplet Loss

In recent years, palm vein recognition has obtained significant attention as its uniqueness, stable features, and high recognition rate. Although state-of-art deep learning methods can outperform several research domains, the lack of sufficiently large data for vein-based biometric recognition can suffer from generalization problems and degrades the model accuracy. Our approach trained Generative Adversarial Nets (GAN) with triplet loss for classification as an additional task. Lately, triplet networks are widely applied as it learns the latent space representation between neighbors and performs significantly higher accuracy even for insufficient data size. Moreover, in practical application, the quality of acquired vein images is low due to external factors and affects the recognition accuracy. To overcome this problem, we propose a CNN-based Encoder-Decoder network for vein segmentation to utilize the accuracy performance. Jerman enhancement filter is applied to enhance the vein ROI images for labeling the ground truth mask images for training the Encoder-Decoder network.