Efficient Single Image De-raining Using Multi-scale Depthwise Separable Dilated Convolution
Xiaojun Bi, Zheng Chen, Xiali Li, Jianyu Yue · 2021 7th International Conference on Computer and Communications (ICCC) · 2021
Single image de-raining based on deep networks has achieved significant progress, which is attributing to complex network structures and the huge number of learnable parameters. However, these complex networks are not suitable for mobile terminals or embedded devices since they are generally resource-constrained. To this end, we propose a novel Efficient De-raining Network (Eff-DerainNet) by introducing multi-scale depthwise separable dilated convolution. Firstly, we propose two novel lightweight modules to construct our network, i.e., MDSD block and MDSD-ConvGRU. The former could extract multi-scale features when parsing the rainy images and the latter is committed to integrating the deep features generated by the whole network. Secondly, we regard the de-raining problem as a multi-stage task and introducing a mature recurrent learning strategy to improve the de-raining performance. Last, we discuss two different supervision forms with four loss functions to maximize the representation ability of our network. Our Eff-DerainNet could achieve competitive performance compared with the latest state-of-the-art deep-learning models, i.e., DRDNet or RCDNet, and bring 20.3- or 12.3-times compression in terms of memory cost. Extensive experiments verify the efficiency and effectiveness of our method in both quantitative assessments and visual quality.