Dual parallel multi-scale residual overlay network for single-image rain removal

Rain not only degrades the perceptual image quality, but also destroys the visibility of the scene, which affects the computer vision algorithms’ ability to capture images outdoors. In response to the above problems, we propose a dual parallel multi-scale residual overlay network (DMRO-Net) for single-image rain removal. First, we combine two parallel branches with different functions. It not only increases the width of the network but also ensures that each layer can obtain rich rain stripe features. Second, the upper branch uses a dual parallel multi-scale residual dense network (DMRD-Net), which can accurately extract rain streak features of different directions, sizes, and densities in rain images. The lower branch uses the residual U-Net (RU-Net), which can expand the receptive domain to obtain more contextual information while avoiding gradient disappearance. Finally, the rain streak features obtained from the two branches are superimposed to obtain the complete rain streak information for the deraining task. We conducted tests on synthetic and real datasets to obtain subjective results and objective evaluations. Experimental results show that our algorithm is more robust and outperforms eleven other different classical algorithms in removing rain streaks for rain image deraining tasks with different intensities of rain streaks. Our algorithm produces a visually clearer deraining image and the image visibility enhancement is effective for computer vision applications (Google Vision API).