Super-resolution image reconstruction (SRIR) endeavors to restore high-resolution (HR) images with enhanced detail from corresponding low-resolution (LR) inputs. With the rapid development of deep learning, integrating deep learning methods provides new solutions for the super-resolution (SR) field. This paper first reviews the background and significance of SR, development process, and the technical value of applying deep learning to SR. Next, SR methods based on deep learning are categorized according to different network types, with a focus on analyzing and comparing the applications of Convolutional Neural Networks (CNNs), Residual Networks (ResNet), Generative Adversarial Networks (GANs), and Diffusion Models in SR. The paper also introduces key evaluation metrics and problem-solving strategies, followed by a performance comparison of mainstream methods on publicly available datasets. Finally, a summary of SR algorithms based on deep learning is provided, along with an outlook on future development trends in the field and explore the possible next research directions in the field of image super-resolution.
Read full abstract