Global attention network with rain prior for real time single image deraining

Poor visibility caused by rainy image can have a negative impact on the performance of computer vision applications. While several image deraining algorithms have been popularly adopted, most of them suffer from two main limitations: (1) they cannot well handle real and complex rain scenes by only focusing on one type of rain in images (e.g. raindrops or rain streaks) whereas the reality often coexists with both types, (2) they face significant difficulties in practical application because of ignoring the speed of inference. To address the above problems, we propose a global attention network (GANet) that can quickly and effectively separate rain streaks and raindrops. Inspired by the fact that rain in images often appears white, we leverage this prior to obtain an initial rain-free background image to guide neural network-based image deraining. Moreover, a new global attention block (GAB) is designed to simultaneously extract the rain features from spatial and channel dimensions. By cascading multiple GABs, the proposed method can effectively obtain the features of rain streaks and raindrops and progressively separates the rain-free image. Furthermore, owing to the nonlinear properties of GAB, the activation functions are omitted, which can speed up the inference time. And the depth-wise and point-wise convolutions are employed to promote computation efficiency as well. Extensive experiments on raindrop and rain streak datasets demonstrate that our method outperforms state-of-the-art methods, achieving up to 37.53 dB PSNR on Rain100L with an inference speed of 39 FPS, which is 2–30 times faster than competitors.