EfficientDeRain: Learning Pixel-wise Dilation Filtering for High-Efficiency Single-Image Deraining

• DOI: 10.1609/aaai.v35i2.16239
• Publication Date: 2022-09-08T18:03:46Z
• Authors: Qing Guo; Jingyang Sun; Felix Juefei-Xu; Lei Ma; Xiaofei Xie; Wei Feng; Yang Liu; Jianjun Zhao
• Keywords: Dilation (metric space); Code (set theory); Kernel (algebra)

Single-image deraining is rather challenging due to the unknown rain model. Existing methods often make specific assumptions of model, which can hardly cover many diverse circumstances in real world, compelling them employ complex optimization or progressive refinement. This, however, significantly affects these methods' efficiency and effectiveness for efficiency-critical applications. To fill this gap, paper, we regard single-image as a general image-enhancing problem originally propose model-free method, i.e., EfficientDeRain, able process rainy image within 10 ms (i.e., around 6 on average), over 80 times faster than state-of-the-art method RCDNet), while achieving similar de-rain effects. We first novel pixel-wise dilation filtering. In particular, filtered with kernels estimated from kernel prediction network, by suitable multi-scale each pixel be efficiently predicted. Then, eliminate gap between synthetic data, further an effective data augmentation RainMix) that helps train network handling images. perform comprehensive evaluation both real-world datasets demonstrate our method. release model code https://github.com/tsingqguo/efficientderain.git.