A Two-Branch Feature Modulated Network for Single Image Deraining

Rain removal from a single image has been attracting much research attention in computer vision. However, most exiting deraining methods explore the function maps from the input rain images to the clear deraining images, which ignore the information of rain streaks for modulating the feature maps of the deraining image. In this paper, we proposed a two-branch feature modulated network (TBFMNet) for exploring the clear deraining image and rain streak, respectively. During the network, the feature maps of the image branch are modulated by the attention feature maps from the rain-streak branch. Meanwhile, attention feature maps are extracted based on spatial attention unit (SAU) and channel attention unit (CAU), which can explore the essential spatial position and channel weight of rain streak. The extensive experiments demonstrate that the proposed method significantly boosts performance for single image deraining, including light rain scene and heavy rain scene.