The method in this chapter is a fur

The method in this chapter is a further improvement of the single-image rain removal algorithm based on the residual multi-scale model described in the third chapter. In the method in Chapter 3, rain lines of different sizes are removed at different image resolutions. In this way, the heavy rain in the original image is easier to remove in the low resolution image, and the lighter rain is easier to remove in the higher resolution image than to remove the rain directly on the original resolution image. Therefore, this method avoids categorizing the size of rainlines in display, and reduces the difficulty of directly removing rainlines of different sizes in the original resolution image. In addition, by making full use of the residuals between the up-sampling results of each stage of rain removal and the resolution rain map of the next stage, the background information lost in the next stage and the location and area information contaminated by rain are provided. , Through channel merging with the input rain map of the next stage and input to the rain removal network to guide the rain removal work in the next stage. On the one hand, in this method, as the image resolution increases and more and more background details need to be restored, the parameters of the rain removal network in the three stages from low-resolution processing to high-resolution processing gradually increase ; On the other hand, the rain removal network of three different stages is trained step by step, and the training efficiency is relatively low.