Abstract
Deraining is an essential preprocess for many computer vision tasks, e.g., vision-based autonomous driving. The existing methods usually depend on the prior information or specified network structures and correspondingly suffer from high computational costs. To enhance the performance to meet the real-time requirement of autonomous driving, we propose a novel end-to-end recurrent multi-level residual learning deraining network featured with the global attention mechanism and residual network architecture. In the proposed Recurrent Multi-level Residual and Global Attention Network (RMRGN in short), we employ a recurrent stage scheme to gradually utilize global contextual information and image details to remove the rain streaks progressively. The global-attention mechanism enables us to focus on the meaningful context in every recurrent stage, which further benefits the network to distinguish the rain streaks and the rain-free images. By exploring the attention information, we further propose a deep multi-level residual learning network to eliminate rain streaks in a single image. Comprehensive experimental results demonstrate that RMRGN performs favorably against the state-of-the-art methods for removing rain streaks.
