Multi-scale Fusion Residual Network For Single Image Rain Removal

Multi-scale Fusion Residual Network For Single Image Rain Removal

Deep learning has been widely used in single image rain removal and demonstrated favorable universality. However, it is still challenging to remove rain streaks, especially in the nightscape rain map which exists heavy rain and rain streak accumulation. To solve this problem, a single image nightscape rain removal algorithm based on Multi-scale Fusion Residual Network is proposed in this paper. Firstly, based on the motion blur model, evenly distributed rain streaks are generated and the dataset is reconstructed to solve the lack of nightscape rain map datasets. Secondly, according to the characteristics of the night rain map, multi-scale residual blocks are drawn on to reuse and propagate the feature, so as to exploit the rain streaks details representation. Meanwhile, the linear sequential connection structure of multi-scale residual blocks is changed to a u-shaped codec structure, which tackles the problem that features cannot be extracted effectively due to insufficient scale. Finally, the features of different scales are combined with the global self-attention mechanism to get different rain streak components, then a cleaner restored image is obtained. The quantitative and qualitative results show that, compared to the existing algorithms, the proposed algorithm can effectively remove rain streaks while retaining detailed information and ensuring the integrity of image information.