Single image rain removal via multi-module deep grid network

摘要

Rain streaks severely degenerate the performances of image/video processing tasks, therefore effective methods for removing rain streaks are required for a wide range of practical applications. In this paper, we introduce an end-to-end deep network, called GridDerainNet, to remove rain streaks within single image under different conditions. The architecture of GridDerainNet consists of three modules: pre-processing, multi-scale attentive module and post-processing. The pre-processing module can effectively generate several variants of the given rainy image, in order to extract more key features from the input. The multi-scale attentive module implements a novel attention mechanism, which allows more flexible information exchange and aggregation, taking full use of diversities of a given image. In the end, post-processing module furthers to reduce residual artifacts after previous two steps. Quantitative and qualitative experimental results demonstrate that the proposed algorithm outperforms several state-of-the-art methods on both synthetic and real-world images.