Single Image Deraining Based on Denoising Diffusion Implicit Models

To address the challenges of prolonged sampling time and the limitation to process specific image sizes in current image deraining research, we propose a novel single image deraining approach based on denoising diffusion implicit models. The method adopts a variant of the U-Net fully convolutional network architecture based on SR3, concatenating rainy images of arbi-trary sizes and noise images along the channel dimension to iteratively obtain high-resolution rain-free images. Specifically, we in-corporate residual blocks from BigGAN-deep to replace traditional residual blocks, thereby deepening the network to enhance model generalization. Additionally, we employ smooth noise esti-mation via image overlapping block processing to improve image fidelity. Introducing deterministic acceleration sampling, we expe-dite the sampling process by uniformly selecting time steps. Comprehensive testing on synthetic and real datasets demonstrates that the proposed method achieves deraining for images of any size, yielding significant improvements in peak signal-to-noise ra-tio and structural similarity. The generated images preserve more complete details, showing superior deraining efficacy.