Progressive Multi-scale Deraining Network
Thatikonda Ragini, Kodali Prakash · 2022 IEEE International Symposium on Smart Electronic Systems (iSES) · 2022
Removing rain streaks from the captured single rainy images plays a dominant role in many computer Vision (CV) applications. Since, many existing deraining methods ignores structural content, introduces artifacts and fails to remove heavy rain streaks completely. To address this issue, a novel PMSDNet was proposed. We introduced an Encoder-Decoder network (UNet) structure in the earlier stagesto extract multi-scale contextual information, whereasfinal stage operates with the Image Original Resolution Network (IORNet) and generates spatially derained outputs accurately. Next a “Feature Fusion Cross Stage” (FFCS) module was inserted between two Encoder-Decoders, Encoder-Decoder and IORNet modules and which makes feature refinements and then propagates to next stage for feature aggregation. At-last an Attention Module (AM) was introduced in between every two stages and generates attention map by suppressing the less informative features at current stage and allows only to propagate useful features to the next stage. Experimental results on synthetic datasets shows that the proposed network achieves state-of-the-art (SOTA) results.