The advent of diffusion models in image super-resolution (SR) has marked a transformative era in generating high-fidelity, high-resolution images from lower-resolution counterparts. However, traditional approaches to diffusion-based SR have grappled with significant challenges,noticeable degradation of image quality, producing outputs that lack sharpness and detail. Thus, our research presents a novel plug-and-play module for any diffusion-based image super-resolution (SR) method, which improves image details by incorporating an edge detection algorithm into the reverse diffusion process.During each denoising step in the reverse diffusion process, we utilize edge map to guide the the image reconstruction towards preserving and emphasizing these key details. This targeted guidance helps in enhancing the sharper and more defined edges that are pivotal for high-quality HR images.
Method overview: The method advances diffusion-based image super-resolution by implementing edge guidance selectively in the later stages of processing to optimize clarity and edge detail. Edge guidance is initiated at step S=0.5T, where T is the total number of diffusion steps. This ensures the guidance is effective when image features become distinguishable, enhancing the precision of the process. The method transforms latent representations into spatial images using a VQ-GAN decode at the chosen step S, followed by applying the PiDiNet architecture for edge detection. PiDiNet processes the image to output a predicted edge map \( \tilde{e}_s \), which is rigorously evaluated against the actual edge map \( e \) using the loss function \( L(\tilde{e}_s, e) = \| \tilde{e}_s - e \|^2 \). This quantifies discrepancies and adjusts the process using the anti-gradient of the loss, \( \tilde{z}_{S-1} = z_S - \alpha \nabla_{z_S} L(\tilde{e}_s, e) \). These steps ensure that EdgeSR not only enhances resolution but also significantly improves edge definition in super-resolved images.
| Methods | PSNR ↑ | SSIM ↑ | LPIPS ↑ | CLIPIQA | ESSIM ↑ |
|---|---|---|---|---|---|
| ResShift | 31.4 | 0.76 | 0.069 | 0.936 | 0.73 |
| EdgeSR | 31.66 | 0.77 | 0.085 | 0.936 | 0.76 |
Quantitative comparison of ResShift and EdgeSR methods on images taken from test set RealSet65, which is consisted of 35 LR images widely used in recent literatures and 30 images were obtained from the internet. The results show that EdgeSR outperforms ResShift in terms of PSNR, SSIM, LPIPS, and ESSIM, indicating superior image quality and edge fidelity.
Comparative visual results showcasing the improvements of our EdgeSR method over the ResShift method.
If you use our work in your research, please cite our publication:
@article{EdgeSR,
title={EdgeSR:Image Super-Resolution Using Edge-Guided Diffusion Models},
author={Panosyan, Armine and Khachatryan, Levon},
year={2024}
}