ENHANCEMENT AND RESTORATION OF IMAGES UNDER ADVERSE VISUAL CONDITIONS USING DEEP LEARNING TECHNIQUES

Abstract

The rapid development of deep learning has revolutionized the field of image enhancement and restoration, particularly for images captured under challenging conditions such as low-light and underwater environments. This thesis investigates the efficacy of advanced deep neural network architectures in improving image quality, visibility, and structural fidelity in scenarios where traditional methods often fall short. Leveraging state-of-the-art models-including architectures with edge-aware modules, attention mechanisms, and transformer-based context modeling-this research demonstrates significant improvements in both quantitative metrics (such as PSNR, SSIM, and LPIPS) and qualitative visual outcomes. Experiments conducted on benchmark datasets, including LOLv1, LOLv2, SID, LSUI, EUVP, and UFO-120, reveal that the proposed frameworks achieve high restoration accuracy and efficiency, with real-time processing capabilities suitable for deployment in resource-constrained environments. The results show substantial gains over traditional and contemporary baselines, confirming the models’ robustness and adaptability across diverse real- world conditions. This study further addresses practical considerations such as computational demands, generalization, and the integration of these methods into applications ranging from surveillance and autonomous navigation to marine exploration and medical imaging. The findings highlight the transformative potential of deep learning in advancing image enhancement technology, offering scalable and effective solutions that benefit a broad spectrum of scientific, industrial, and societal domains.