Abstract: A deep feature fusion method was proposed for low-light images rain streak removal in digital twin applications. First, the dynamic line convolution（DLConv） was created to learn the prior knowledge of rain streaks, enhancing the ability to extract rain streak features to address the artifact problem. Second, by combining prior knowledge with the coordinate attention（CA） mechanism, the U-shaped model was improved to generate deep features of rain streaks, thereby addressing the issue of detail loss. Finally, the brightness information of the images was utilized to further refine and enhance the deep features and obtain rain-free images at the current stage, where the features of the rainfree images were fused between the stages, and multiscale features were transferred through cross-stage feature fusion, achieving rain streak removal for low-light images. Experimental results on the latest low-light rainy image dataset showed that the proposed method achieved peak signal-to-noise ratio（PSNR）、structural similarity（SSIM）、learned perceptual image patch similarity（LPIPS） and natural image quality evaluator（NIQE） values of 38.522 9 dB, 0.974 5, 0.061 9 and 4.207 2, respectively, outperforming existing mainstream methods. This demonstrated the effectiveness of the proposed approach in alleviating artifacts and detail loss, thereby significantly improving the usability of low-light images.

Key words: digital twin, low-light image, image rain streak removal, deep feature, coordinate attention（CA） mechanism