Single Image De-raining via Densely Connected Pyramid Network

Abstract

Images captured in rainy outdoor scenes often suffer from poor visual quality due to the appearance of rain streaks in the image. These degraded images drastically affect the performance of many practical vision systems. Hence, image de-raining is urgently required to be addressed. In this paper, we propose a new deep network for single image deraining, called Densely Connected Pyramid Network (DCP-Net). As the dense connection can maximize the information flow along features from different layers and the multi-scale strategy has been successfully applied in many computer vision tasks, we combine these two benefits to design a Densely Connected Pyramid Block (DCPB) as the basic de-raining unit, called DCPB-Unit and introduce a dense connection architecture to connect several DCPB-Units, which strengthens the feature propagation between the DCPB-Units and improves the effectiveness of learning. The proposed network efficiently makes use of the features from multiple layers and learns the rain streaks with different scales and shapes. Specifically, benefitting from computing the rain-free image by subtraction operation in the feature domain rather than in the image domain, the network is able to generate the high-quality rain-free image. Both quantitative and qualitative experimental results demonstrate the proposed method performs favorably against the state-of-the-art de-raining methods.