A Super-Resolution Approach for Image Resizing of Infant Fingerprints With Vision Transformers

Digital biometrics is one of the most widespread and reliable means of recognizing human identity. Despite being well-established for adults, digital biometrics for newborns constitutes a considerable open challenge in the literature. The very small size of the fingers exposes subtle features that can neither be visually inspected nor computationally captured. This leads conventional biometric matchers, which generally rely on the distance between ridges, not to work as expected for newborns. To some extent, this problem can be mitigated by increasing the scanner’s resolution, which is an expensive strategy. Besides, the saturation limit for resolution is unknown beforehand. Another alternative is to resize images using interpolation, but this can introduce unexpected noises or highlight preexisting ones. This paper proposes an approach for image resizing based on Super-Resolution networks that train Vision Transformer (ViT) how to handle newborn fingerprints, avoiding the introduction of noises while keeping and improving important details for biometric recognition. For biometric identification and verification, results show an increase in the order of 10% for resized images in comparison with traditional resizing algorithms.