AN ENHANCED HYBRID DEEP RESIDUAL AND MULTI-SCALE FEATURE NETWORK FOR HIGH-FIDELITY IMAGE SUPER-RESOLUTION

Abstract

Super-resolution of remote sensing and aerial images is a critical preprocessing step for improving visual interpretation of images and downstream computer vision tasks. Traditional interpolation-based methods often fail to recover high-frequency details from the image as well as degrade the quality of image, while single-scale deep learning models struggle to capture both fine textures and global contextual information. To overcome this problem, this paper proposes a Hybrid Very Deep Super-Resolution–Multi-Scale Feature Network (VDSR-MSFN) for single-image super-resolution. The proposed framework incorporate deep residual learning through multi-scale feature extraction to efficiently reconstruct high-resolution images from low-resolution inputs. Initially, bicubic-upsampled luminance images are processed through a VDSR-based deep residual stack to learn local high-frequency structures. Consequently, multi-scale feature extraction blocks comprise parallel 3×3, 5×5, and dilated 3×3 convolutional branches are employed to capture both local and global spatial information. The extracted features are concatenated and fused using 1×1 convolutions, followed by local and global skip connections to preserve low-frequency components and improve gradient flow. A reconstruction component predicts the high-frequency residual, which is added to the bicubic input to produce the final super-resolved image. Experimental results demonstrate that the proposed hybrid building achieves superior reconstruction performance in terms of PSNR, SSIM, and MSE.