Frequency-aware deep networks and patch-wise generative adversarial training for single image super-resolution in wavelet domain

With the increasing demand for high-resolution displays, image super-resolution has become essential for enhancing visual quality, especially in the era of 4K and 8K devices. Single image super-resolution (SISR) plays a vital role in numerous computer vision applications because of its ability to enhance image resolution while preserving important details. Recently, deep learning-based super-resolution methods, including methods based on generative adversarial networks (GANs) and transformers, have become mainstream. However, the existing methods face the challenge of preserving high-frequency details while maintaining overall image quality. They often struggle to reconstruct realistic and undistorted high-frequency information. To address this challenge, we propose the frequency-aware perceptual wavelet domain super-resolution (FA-PWSR) model. FA-PWSR deals with images’ low-frequency and high-frequency components separately and in parallel. FA-PWSR employs a divide-and-conquer strategy, utilizing stationary wavelet transform (SWT) to decompose images into low and high-frequency sub-images. Specialized subnetworks are designed to process each component, ensuring targeted optimization for different frequency bands. The reconstructed sub-images are subsequently integrated using the inverse stationary wavelet transform (ISWT) to generate the final SR image. FA-PWSR achieves superior performance over state-of-the-art methods in both fidelity (PSNR and SSIM) and perceptual metrics (PI and LPIPS) across various datasets including Set5, Set14, Urban100, and BSD100. Even in the challenging Urban100 dataset, FA-PWSR achieves a 1.754 dB improvement in PSNR and a 0.049 increase in SSIM compared to ESRGAN. Furthermore, our model significantly enhances perceptual quality, achieving a 12 % reduction in LPIPS on the same dataset. Moreover, visual comparisons confirm that FA-PWSR effectively preserves fine details, enhances edges, and noticeably reduces artifacts, leading to more realistic and high-fidelity super-resolved images.