Abstract <p>For ultrahigh definition images, we propose an end-to-end single image deraining network. The input image is divided into several patches that share encoder and decoder parameters. The encoder and decoder are designed by a linear Transformer to reduce computational complexity. Then we fuse the features of each subimage obtained from the encoder and employ a vision Mamba-based bottleneck to extract relationships from the fused features. Through this creative design, we avoid extremely high memory demands while enabling the learning of correlations between subimages. Our method outperforms state-of-the-art methods, as evidenced by experimental results on the ultrahigh definition dataset.</p>