Finger-vein recognition using a novel enhancement method with convolutional neural network

Finger vein biometric technology has gained a lot of popularity over recent years. This is primarily due to the increased security and reliability level that comes with its non-intrusive nature. Non-intrusiveness became inevitable due to the pandemic of COVID-19. This paper introduces a unique and lightweight image enhancement method for person identification using Convolutional Neural Networks (CNN). As pre-processing steps, Contrast Limited Adaptive Histogram Equalization (CLAHE) followed by gamma correction is applied. Afterward, the image is sharpened and then passed through the median filter. These steps are followed by applying power law and contrast adjustment. As a final step, CLAHE is used yet again to bring out the enhanced vascular structure. The method was appraised using the four different openly accessible databases. These are regarded as the most challenging available finger vein database-s by many researchers. For recognition purposes, CNN was used with transfer learning. Transfer learning is implemented by modifying the 13 convolutional layers of VGG-16. The proposed model architecture also includes five max-pooling layers, one ReLU, and one Softmax layer. It is observed that with transfer learning, the accuracy could have reached up to 99% on finger-vein recognition on the experimented dataset, thus proved to be a highly accurate approach for finger vein recognition.