Abstract: To address the excessive rain removal and poor generalization of images, this paper proposes a single-image de-raining method by improving diffusion model. The data becomes Gaussian distribution by adding Gaussian noise to the forward process. The dual input information channels of the residual module are designed and the ECA (Efficient Channel Attention) channel attention mechanism module is added to build a noise estimation network. Thus, a global average pooling without reducing the dimension is achieved and the local cross-channel interaction information is captured. The model network is employed to reverse sampling, predict the noise as a rain mark and remove it, and thus achieve image de-raining. By employing simulated raindrop datasets and the Rain100 dataset, comparative experiments are conducted to compare our improved diffusion model with other four algorithms. The experimental results demonstrate our improved diffusion model effectively removes rain streaks, with peak signal-to-noise ratios of 30.328 5 for raindrops and 34.896 5 for rain lines, and structural similarities of 0.927 1 and 0.962 0 respectively. A real rain image dataset is built, and the YOLOv7 algorithm is employed to perform vehicle detection on the rain-removed images. Our results show the improved diffusion model for rain removal effectively enhances the confidence of vehicle detection, further confirming it has outstanding de-raining performance and generalization capability.