Enhanced EDSR-based Deep Neural Model for High-Fidelity Image Super-Resolution and Noise Reduction

Abstract

High-fidelity image super-resolution and effective noise reduction remain persistent challenges in computer vision, where maintaining structural accuracy while enhancing perceptual quality is essential. Conventional methods, including GAN-based architectures such as ESRGAN and Real-ESRGAN, have demonstrated promising results but still encounter issues like texture distortion, visual artifacts, and limited representation of complex image dependencies. To overcome these limitations, this study introduces an enhanced EDSR-based deep neural model tailored for image super-resolution and noise reduction. The proposed framework builds upon the strengths of residual learning and deep convolutional layers, refined with optimization strategies to better capture fine-grained details and suppress noise across varying degradation conditions. Comprehensive experiments are conducted on benchmark datasets including DIV2K, Set5, and Urban100, using full-reference metrics (PSNR, SSIM) along with perceptual quality assessments (LPIPS, NIQE). Experimental findings confirm that the enhanced EDSR model achieves superior performance compared to traditional CNN and GAN-based approaches, delivering sharper reconstructions, improved texture fidelity, and effective noise suppression. Additionally, the model demonstrates strong generalization to real-world low-quality images, underscoring its potential in diverse application domains such as medical imaging, remote sensing, and digital photography. This research advances deep neural solutions for image enhancement, presenting a robust and scalable framework for high-quality super-resolution and denoising tasks.