Deep Learning-Based Small Sample Body Vein Recognition and Its Implementation

Abstract In scenarios like forensic investigations, only RGB images of body parts, especially limbs, are available from surveillance. Traditional biometric methods like face, fingerprints, and iris scans are not applicable in such cases. Additionally, hand veins, typically captured using NIR imaging, are not accessible either. To address this challenge, body vein recognition technology has been introduced. Different techniques have been proposed to uncover and extract vein patterns from RGB images. However, handcrafted features are sensitive to image noise, and deep learning methods require extensive training data, which is lacking in forensic investigations. In this study, we presented a small-sample body vein recognition approach based on deep learning. A data augmentation method utilizing Gabor filtering and vein line traction was introduced to enhance generalization performance and tackle the limitation of training samples. A new model combining improved DenseNet and Arcface loss was developed for body vein recognition. Furthermore, a feature retrieval scheme based on binary hash coding was introduced to optimize recognition accuracy and retrieval efficiency. Experimental results using images from 250 forearms and 230 thighs demonstrated the superiority of our small-sample body vein recognition method over other matching techniques. Ablation studies confirmed the effectiveness of the data augmentation and feature retrieval schemes. Finally, the model was deployed on a portable system, leading to the implementation of a contactless body vein recognition system.