An Enhanced Very-Deep Super-Resolution (VDSR) Neural Network for Single Image Super-Resolution using Luminance and Chrominance Channels

Mazlinda  Ibrahim; Muhammad Anas  Zulkeffly; Mat Kamil  Awang; Hoo  Yann Seong; Nurul Kamilah  Mat Kamil

Authors

Mazlinda Ibrahim Centre for Defence Foundation Studies, National Defence University of Malaysia, 57000 Kuala Lumpur, Malaysia
Muhammad Anas Zulkeffly Department of Electrical and Electronic Engineering, Faculty of Engineering, National Defence University of Malaysia, 57000 Kuala Lumpur, Malaysia
Mat Kamil Awang Department of Electrical and Electronic Engineering, Faculty of Engineering, National Defence University of Malaysia, 57000 Kuala Lumpur, Malaysia
Hoo Yann Seong Centre for Defence Foundation Studies, National Defence University of Malaysia, 57000 Kuala Lumpur, Malaysia
Nurul Kamilah Mat Kamil Faculty of Electrical Engineering and Information Technology, RWTH Aachen, Aachen, Germany

Keywords:

Super Resolution, Neural Network, Image Restoration, Image Reconstruction, Perceptual Quality

Abstract

Single Image Super-Resolution (SISR) is a fundamental problem in computer vision, aiming to enhance the resolution of a low-resolution image while preserving its details and structure. The Very-Deep Super-Resolution (VDSR) network has demonstrated significant success in SISR tasks due to its depth and residual learning framework. However, its performance can be further improved by utilizing additional image information. In this paper, we propose an enhanced VDSR network that integrates luminance and chrominance channels as supplementary inputs. By separately processing the luminance and chrominance components, the network learns complementary features, enabling more accurate reconstruction of high-frequency details and textures. This approach aligns with the human visual system's heightened sensitivity to brightness changes, resulting in super-resolved images with superior perceptual quality. Experimental results on available datasets confirm that the proposed method outperforms the original VDSR, achieving notable improvements in natural image quality evaluator, blind image spatial quality evaluator and perception-based image quality evaluator. The decrease in the performance metrics is up to 2%. This study underscores the effectiveness of incorporating luminance and chrominance information in SISR tasks, paving the way for more accurate and visually appealing image reconstructions.