Image De-Raining Method Based on PReNet and Diffusion Model

Existing image de-raining methods often suffer from incomplete rain streak removal and the disappearance of background information during the de-raining process. This paper suggests a novel de-raining approach based on the composition of PReNet and diffusion models. By improving the PReNet network structure, incorporating a CA Layer attention mechanism module, and enhancing the recursive layers, the feature extraction capabilities are strengthened. The noise estimation network in the diffusion model is replaced with the improved PReNet structure to leverage its denoising abilities for effective rain removal. Network training is conducted on synthetic datasets, and both qualitative and quantitative results are obtained. Quantitatively, on the Rain100H and Rain100L datasets, the proposed method achieves average PSNR values of $\mathbf{2 6 . 5 1} \mathbf{~ d B}$ and 33.24 dB, and SSIM values of 0.8310 and 0.9633 respectively, outperforming current state-of-the-art methods. Qualitatively, the suggested approach retains more image details shows good generalization to various kinds of rain streaks, and avoids image distortion.