基于条件扩散隐式模型单幅图像去雨

Rain Removal for a Single Image Based on ConditionalDiffusion Implicit Models

雨水天气会对图像造成干扰并增加图像处理的难度。为消除雨水对图像造成的影响,提出一种基于条件扩散隐式模型的图像去雨方法。该方法采用基于SR3的全卷积网络架构,使用U-Net结构的变体,并用BigGAN的残差块替换了传统的残差块,去掉自注意力机制、位置编码和群组归一化,实现了条件扩散模型支持任意大小图像的输入,且不受图像分辨率的影响。同时,引入确定性加速采样,用子序列时间步来加速生成过程,提高图像恢复速率。通过对图像进行重叠分块处理,将子块分次调入内存处理,减少资源消耗,提高算法的适用性,使用平滑噪声估计引导去噪过程,使生成图像获得更高的保真度。在合成数据集和真实数据集上进行测试,定性和定量结果表明,该方法在峰值信噪比和结构相似性方面均有提升,图像细节信息保留更加完全且去雨后的视觉效果更佳。

Rainy weather can interfere with images and increase the difficulty of image processing.To eliminate the influence of rain on images,we propose an image rain removal method based on the conditional diffusion implicit model.It adopts a fully convolutional network architecture based on SR3,using a variant of the U-Net structure,and replaces the traditional residual blocks with BigGAN's residual blocks,removing self-attention mechanism,positional encoding,and group normalization.This implementation enables the conditional diffusion model to support input of images of any size,without being affected by the image resolution.In addition,the proposed method introduces deterministic accelerated sampling and uses subsequence time steps to speed up the generation process and improve the image restoration rate.By processing the image in overlapping blocks,the sub-blocks are processed and called into memory in stages,reducing resource consumption and improving the algorithm's applicability.The method uses smooth noise estimation to guide the denoising process,achieving higher image fidelity for the generated images.Tests on synthetic and real datasets show that the proposed method improves the peak signal-to-noise ratio and structural similarity of images,while preserving more complete details and achieving better visual effects after rain removal.