Density-aware Single Image De-raining using a Multi-stream Dense Network.

Single image rain streak removal is an extremely challenging problem due tothe presence of non-uniform rain densities in images. We present a noveldensity-aware multi-stream densely connected convolutional neural network-basedalgorithm, called DID-MDN, for joint rain density estimation and de-raining.The proposed method enables the network itself to automatically determine therain-density information and then efficiently remove the correspondingrain-streaks guided by the estimated rain-density label. To better characterizerain-streaks with different scales and shapes, a multi-stream densely connectedde-raining network is proposed which efficiently leverages features fromdifferent scales. Furthermore, a new dataset containing images withrain-density labels is created and used to train the proposed density-awarenetwork. Extensive experiments on synthetic and real datasets demonstrate thatthe proposed method achieves significant improvements over the recentstate-of-the-art methods. In addition, an ablation study is performed todemonstrate the improvements obtained by different modules in the proposedmethod. Code can be found at: https://github.com/hezhangsprinter