Vascular Enhancement Analysis in Lightweight Deep Feature Space

Finger-vein has been used for automatic personal recognition systems for over a decade. Its size and characteristics against presentation attacks make it suitable for various commercial and governmental applications. As with other biometric traits, the majority of finger-vein recognition methods employ hand-crafted feature for image classification. Lately, there are some convolutional neural network (CNN) models are designed for finger-vein identification and verification. Whereas, the CNN-based approaches tend to set up increasing layers and more parameters, which incur equipment memory issue and algorithm execution speed issue. Besides, many researchers have made great efforts to enhance hand-crafted feature-based method by using image enhancement algorithms. Hence, it is significant to explore high-performance finger-vein recognition without image enhancement in deep learning world. In this work, two different dimension issues are investigated: lightweight CNN model, the effect of image enhancement on finger-vein biometrics in deep feature space. Transfer learning plays an important role in extracting informative and representative features by information transfer from the pre-trained network to a novice via fine-tuning. The experimental results on four benchmark databases HKPU, FV-USM, SDUMLA and UTFVP demonstrate the proposed lightweight model ‘LightFVN’ outperforms the existing published competition winners. Furthermore, image enhancement turns out not to be essential for CNN-based finger-vein biometrics.