Improving Ear Recognition with Super-resolution

Super-resolution (SR) techniques are advanced image enhancement methods that increase the resolution of an image, proving to be valuable in various applications such as enhancing photographs and boosting person recognition performance. In this study, we focus on Single Image Super Resolution of ear images, a super-resolution approach that generates missing information from a single input image. Early super-resolution attempts relied on interpolation, but these methods had limited success. The advent of deep neural networks has led to significant improvements in SR performance. We assess the performance of the Enhanced Deep Residual Network (EDSR) and Shifted Windows Transformer Network (SwinIR) for ear image super-resolution and extend the study to recognition experiments using super-resolution images. Utilizing the UERC dataset, which contains 16,665 ear images with varying sizes, shapes, and orientations, we trained four models: two on EDSR and two on SwinIR networks, with scaling factors of two and four, respectively. Recognition experiments were employed to evaluate the two distinct model designs. The EDSR model demonstrates superior performance in terms of Rank-1 recognition. Recognition experiments using super-resolution images were conducted using original images and bicubic interpolated images as a baseline. A regular ResNet was used to showcase applicability of super-resolution for ear recognition in virtually any ear recognition model.