Dense Residual Generative Adversarial Network for Rapid Rain Removal

Abstract

Single-image rain removal always was one of the difficulties in the environment perception task. Usually, it has two paths to solve this problem: data-driven solutions and model-based solutions. Due to the benefits of convenient, learning features automatically and rapidly, data driven solutions has attracted tremendous interests. However, the time consumed per frame is still hard to match the requirement of high real-time performance, especially for high speed unmanned platform. In this article, we propose a fast dense residual generative adversarial network (FDRN), which can remove rain and reduce computation time consumption, the de-raining time of each frame only consumes 0.02s. We enhanced the data of original rainy images, put it into the generator network which is composed of long short-term memory networks (LSTM) and a newly designed dense residual network (DRN). The feature map in generator and discriminator is extracted to calculate the loss function and guide the direction of training. We selected 1500 pairs of synthetic images from existed datasets to train our network. And in order to test our method's de-raining ability realistically, we also selected 147 real-world rainy images from existed datasets. Experiments on both synthetic and real-world rainy images demonstrate that the proposed method can achieve competitive results to some existing methods in performance and effectiveness.