Rain Streak Attention Network for Single Image Deraining
Zhiguo Kang, Wei Liu, Caiwang Zhang, Minghui Li · 2024
Single image deraining methods have been extensively studied for its ability to remarkably improve the performance of computer vision tasks in rainy environments. However, existing methods are rarely able to use the properties of the rain streaks themselves to remove rain, and they often have difficulty correctly identifying the rain streaks and background information when removing complex rain streaks, resulting in the loss of some of the detailed textures. In this paper, we propose a Rain Streak Attention Network, which can make full use of rain features between different color channels to identify rain streaks from complex backgrounds. The basic building blocks of the method are the rain streak extraction block (RSEB) and the detailed feature recovery block (DFRB). Specifically, RSEB aims to separate rainbands from background information by focusing on similar rainfall characteristics in different channels using an attention mechanism. In DFRB, we achieve the modeling of global features by deep-wise dilation convolution, which helps to recover more realistic detailed texture information. Extensive experimental results on the commonly used benchmarks demonstrate that the proposed method achieves favorable performance against state-of-the-art approaches.