Restoring Palmprint Biometrics: A GAN based Hybrid Framework for inpainting and deblurring

Palmprint recognition is a reliable biometric identification technique and restoration is valuable technique for restoring and enhancing images which have been highly distorted either by some missing part or by addition of some noise to the image or the blur images. The conventional denoising algorithms struggle to handle noise whereas Generative Adversarial Networks, GAN are proved to be the efficient generative models that produces promising result and shows remarkable performance in this field. The GAN-based model is effective for denoising low-resolution palmprint images due to its ability to handle noise and retain more orientation information. Several researches are made on the restoration and various GAN models are explored but the challenge was found to be that mostly all the models focus on only single type of restoration, there is still a scope to design a GAN model that works on all types of noises with the comparatively increased efficiency. An intelligent framework/architecture is needed to generalize this complex phenomenon. This research proposes a GAN model that focuses on restoration of image damaged by several noises/factors. This study introduces a novel hybrid GAN-based model that addresses inpainting and deblurring for palmprint repair. It makes use of Transformer blocks, a PatchGAN discriminator, and a U-Net Based generator. The model learns global context and long-range dependence, downsamples, and extracts hierarchical features. The discriminator establishes if the created or real image is authentic or not. 312 subjects' 5,502 palmprint images from the CASIA palmprint library were used to train the model. Various deblurring and painting models are analysed and the proposed model is found to generate better performance. The approach can be used in real-world scenarios because it is end-to-end and doesn't require further noise localization information. Additionally, the model's scalability and processing efficiency are assessed in the article.