Development of novel generative deep learning models for single image super resolution of fingerprint images

Abstract

This thesis presents a comprehensive study on single image super-resolution (SISR) newlinewith a focus on enhancing resolution of low-resolution images using generative deep newlinelearning techniques for fingerprint images. The primary goal of this research is to newlineimprove the recognition, analysis and interpretation of low-resolution images, which newlineare critical in various fields such as biometric identification, forensics, surveillance, newlinemedical imaging and remote sensing. This research introduces several key newlinecontributions, starting with the development of a robust autoencoder incorporating newlineaccelerated skip connections, designed to significantly enhance image quality. newlineAdditionally, a convolutional neural network (CNN) with subpixel convolution newlineoperations was implemented, enabling an 8x resolution enhancement. The work further newlineincludes a computationally efficient generative adversarial network (GAN) newlinearchitecture, optimized for perceptual quality, which generates high-resolution images newlinewhile maintaining a compact model size. Moreover, a lightweight autoencoder based newlinesuper-resolution model was developed, providing a balance between accuracy and newlinecomputational efficiency. All models were built from scratch and extensively tested on newlinefingerprint datasets, demonstrating notable improvements in detail preservation and newlinehigh identification accuracy. The findings have substantial implications across various newlinefields and offer scalable, practical solutions for image resolution enhancement. Future newlinework envisions extending these models to video applications, real-time systems and newlineforensic investigations, further advancing the field of super-resolution and expanding newlineits application scope. newline