Enhancing Finger Vein Recognition With Image Preprocessing Techniques and Deep Learning Models

Abstract

The use of finger veins for biometric authentication is increasingly popular; however, low-quality images present significant challenges that necessitate innovative approaches for accurate identification. This study investigates the effectiveness of various image preprocessing techniques to enhance and analyze finger vein images for biometric recognition. We utilized a finger vein dataset from Kaggle, comprising diverse images captured under controlled conditions. Our preprocessing methods included sharpening images using convolution kernels to improve edge definition, employing thresholding techniques (simple binary, adaptive, and Otsu’s) for effective image segmentation, and applying morphological operations (erosion, dilation, opening, and closing) to refine object shapes and reduce noise. We also implemented edge detection methods, including Sobel, Laplacian, and Canny, to identify significant boundaries within the images. Image resizing was performed using linear, cubic, and area interpolation to assess their effects on image quality. Additionally, various filtering techniques-such as Kalman, median, and Gaussian filters-were applied to reduce noise and enhance image clarity. The dataset comprised 3,816 images from 106 individuals, split into two configurations: 80-10-10 and 70-15-15. We assessed models such as VGG16, VGG19, ResNet101, AlexNet, MobileNet, DenseNet201, and EfficientNet based on metrics including accuracy, precision, recall, F1-score, and training time. VGG16, VGG19, and ResNet101 achieved accuracies of 99.9%, 99.8%, and 99.8%, respectively. Data augmentation techniques generated 76,320 augmented images, significantly improving model performance, especially for the 80-10-10 split. Visualizations through radar and bar charts indicated that VGG16, VGG19, and ResNet101 delivered the highest performance metrics, while DenseNet201 exhibited a slight decline in the 70-15-15 split due to increased test data. Overall, the findings demonstrate the models’ efficacy for reliable finger vein biometric recognition, contributing to advancements in biometric authentication systems.

Keywords

• Finger vein
• biometric recognition
• deep learning
• image preprocessing
• feature extraction
• biometric authentication