Learning Single Image Rain Streak Removal Based on Deep Attention Mechanism
Kuan‐Hua Huang, Li‐Wei Kang · 2023
Bad weather conditions (e.g., rain or hazy) may significantly degrade the visual quality of captured images/videos and the performances of related applications (e.g., outdoor visual surveillance). To solve this problem, this paper presents to learn rain steak removal from a single image. By using the ECNet (Embedding Consistency Network, by Li et al., 2022) as our basis network architecture, a deep encoder-decoder-based network with channel attention and the proposed multi-scale pixel attention module (MSPAM) is presented to single image rain streak removal, i.e., deraining. Together with the "Rain Embedding Consistency" mechanism used in the ECNet, we have shown that the channel attention can be used to enhance the extracted features before being fed into the encoder, and our MSPAM can be embedded into the skip connection between the encoder and the decoder for further boosting the features to achieve better image reconstruction. Experimental results have demonstrated that the proposed framework outperforms the ECNet quantitatively and qualitatively.