Structural-perceptual Image Super Resolution Using Charbonnier-SSIM Loss in an Efficient Sub-pixel Convolutional Network

This paper introduces an image super-resolution framework that combines a subpixel convolutional network with a novel Charbonnier-SSIM hybrid loss to enhance perceptual and structural fidelity. Our hybrid loss encourages better edge preservation and visual consistency. Proposed method has channel adaptivity to perform super-resolution on RGB/Grayscale images. We evaluate the model on the DIV2K and MSAR-1.0 datasets at ×4 scale, achieving superior average PSNR of 28.2dB and improved SSIM scores over ESPCN. Moreover, its compact size is suitable for deployment on edge devices. The network's improved results in more visually pleasing reconstruction with enhanced resolution, which is beneficial in target recognition, terrain mapping, and change detection.