Abstract: Single Image Super-Resolution (SISR) reconstructs high-resolution (HR) images from low-resolution (LR) inputs. Although deep networks achieve strong results, many incur high computational cost and blur fine textures. We present an efficient deep autoencoder for SISR: an encoder extracts compact LR features, and a decoder synthesizes the HR output. To enhance detail recovery, we incorporate skip connections and residual learning within the reconstruction path. Evaluations on the DIV2K dataset show that the proposed model attains high PSNR and SSIM and surpasses several benchmark methods in reconstruction quality. Ablation studies confirm that each enhancement contributes to sharper edges and more consistent texture reproduction across diverse scenes and images. Compared with lightweight architectures such as FSRCNN, our approach uses moderately more parameters and runs slower on CPUs, yet produces clearly improved visual fidelity. These findings indicate that enhanced autoencoder designs are practical for super-resolution, particularly when deployed on GPUs or edge-accelerated hardware.

Keywords: Single Image Super-Resolution, Convolutional Autoencoder, Deep Learning, Image Enhancement, PSNR, SSIM

Pages: 125-132