Deep Scale-Space Mining Network for Single Image Deraining
Images captured by outdoor vision systems can often be affected by rain weather, resulting in severe degradation of the visual quality of the captured images. Therefore, image deraining has attracted attention as urgent and challenging research. Many current data-driven approaches achieve better performance but are limited in recovering image details. This is because these methods do not fully mine the correlation of scale-space, which are beneficial for rain removal. In this paper, we design an end-to-end Deep Scale-space Mining Network (DSM-Net) for single image deraining to solve these problems. The proposed network with multi-scale extraction, concurrent attention distillation, and hierarchical information fusion accurately captures scale-space features and learns richer information for better deraining. For better feature extraction, a Multi-scale Attention Block (MAB) is introduced to obtain multi-scale rain streak features by different dilated convolutions. Besides, a Concurrent Attention Distillation Block (CADB) is developed which combined channel attention and subspace attention to calibrate the image features obtained from multiscale acquisition and hierarchical learning, then eliminate redundant features. Importantly, the overall architecture of DSM-Net is inspired by the HourglassNet and DenseNet, which progressively explores and fuses local and global features at different scales in a hierarchical manner instead of direct concatenation. Extensive experiments on synthetic and real datasets show that the proposed DSM-Net outperforms recent state-of-the-art deraining algorithms in terms of both performance and preservation of image details.