Abstract: The core challenge of the image rain removal task is to be able to remove the rain streak cleanly while preserving the texture details of the background image. With the increase of network depth, most deep CNNS will have problems such as performance saturation and learning decline, and the traditional method of rain removal will cause the loss of background image details due to the increase of network depth. Through the experiment, it is found that although the high level features lost the clear background, but also less rain marks. The lower level features retain rain marks, but the background is clear. In addition, Transformer has a good ability to capture global information and information fusion. In this paper, an image rain removal method based on cross-reasoning is proposed. Transformer adopts the network structure and takes advantage of the high level features of rainless images to pay more attention to the background of low level features and enhance the background repair ability. At the same time, the rain marks of the lower layer are used to give higher attention to the rain marks of the upper layer to enhance the rain removal ability. The empirical results show that the proposed algorithm achieves good results in several rain removal tasks compared with the mainstream algorithms.