Multi-Scale Attention Generative Adversarial Network for Single Image Rain Removal

Abstract

Many severe weather conditions (rain, haze, sandstorm) have a significant impact on the quality of the images; therefore, it severely restricts lots of computer vision tasks such as object recognition, object detection, and object tracking. Although there are plenty of deraining algorithms, single image deraining is relatively rare. In real-world scenarios, rain removal in a single image is a difficult task, and exiting methods often result in poor effects. We present a multi-scale attention generative adversarial network called MSA-GAN for single image rain removal, which applies an attentive generative network using adversarial training. The generative network adopts multi-scale attention mechanisms which use spatial pyramid to capture features from different receptive fields and lead the fine fusion of relevant information at different scales. Extensive experimental results on synthetic and real-world rainy data sets show that our method has better performance than the most state-of-the-art ones. The proposed method also inspires a new research direction of vision task. Our source code is soon to be available.