A Comparative Study of ConvNets based on Finger Vein Biometrics

Gaining extraordinary attention in recent years, Finger Vein Biometrics has an ideal viability with the advantages of being the least susceptible to identity theft and being present inside the body, making it difficult to forge and be affected by any external factors. Although methods of extracting information from the veins are still under investigation, one of the critical stages of biometrics is recognition and matching to the individual under test. It is carried out with techniques of explicit feature extraction, which add to the extra computation and are hence deprecated and unnecessary. Pattern-matching methods using feature extraction techniques add to the time and memory complexity of the models. Instead, one can use machine learning models, like CNNs and ‘commercial ConvNets’ (MobileNET, ResNET, DenseNET, etc.), to obtain an automated feature map for pattern matching with high accuracy. This article presents specially customized CNNs to achieve high matching accuracy for finger vein recognition and compares performance with commercial ConvNET models.