Single image super-resolution based on deep networks and wavelet transform

In the field of image processing, super-resolution (SR) presents significant challenges, notably in recovering high-frequency details lost due to diffraction during image capture. Recent advancements demonstrate that deep learning (DL), particularly generative adversarial network (GAN), effectively addresses single image SR (SISR) tasks. However, GAN-based strategies often introduce fictitious details to enhance image resolution. Moreover, the prevalent assumption that low-resolution (LR) images are merely downsampled bicubically from their high-resolution (HR) counterparts often results in suboptimal performance. To tackle these issues, we introduce a comprehensive framework that employs multi-resolution analysis to learn a residual image super-resolver, enhancing the practicality of the GAN approach by extracting intrinsic information from image degradation. Our methodology assumes that the residuals between subband images and their corresponding input images contain critical information about the actual degradation and downsampling processes. We demonstrate through experiments that learning these residuals facilitates improved detail reconstruction in natural images, allowing our method to surpass current state-of-the-art techniques both quantitatively and qualitatively.