DPCN: Dual Path Convolutional Network for Single Image Deraining

The visual effect of images captured on rainy days is severely degraded, even making some computer vision or multimedia tasks fail to work. Therefore, image rain removal is crucial for these visions and multimedia tasks. However, most existing works cannot strike a good balance between removing rain streaks and restoring the corresponding background detail. To address this problem, this paper proposes an effective dual path convolutional network (DPCN) for single image rain removal. Specifically, we complete the positioning, extraction and separation of rain streaks through multiple dual path units. Firstly, considering the irregularity of the size, density and shape distribution of rain streaks, a pixel-wise attention mechanism is applied to pinpoint the position of rain streaks. Simultaneously, for these rain streaks distributed across regions, we propose a multi scale aggregation method to extract and fuse features at different scales. Further, for some backgrounds with similar texture details as the rain streaks, we introduce a self-calibration operation that separates the rain streaks from these background details by adaptively constructing long-range spatial and internal channel dependencies at each spatial location. By cleverly combining multiple dual path units through a dual path topology, our network obtains rain removal results that are closer to the real background and largely remove rain streaks. The quantitative and qualitative results on synthetic and real datasets show that our proposed DPCN is superior to other state-of-the-art methods.