Single Image Rain Removal via a Simplified Residual Dense Network

The single-image rain removal problem has attracted tremendous interests within the deep learning domains. Although deep learning based de-raining methods outperform many conventional methods, there are still unresolved issues in regards to improving the performance. In this paper, we propose a simplified residual dense network (SRDN) to improve the de-raining performance and cut down the computation time. Inspired by the image processing domain knowledge that a rainy image can be decomposed into a base (low-pass) layer and a detail (high-pass) layer, we train our network by directly learning the residual between the detail layer of rainy images and the detail layer of clean images. It can both significantly reduce the mapping range from input to output and easily employ the image enhancement operation to handle the heavy rain with hazy looks. Instead of designing a deeper network structure to increase the learning ability of network, we propose a simplified dense block to explore more effective information between layers and, hence, reduce the computation time of network. Experiments on both synthetic and real-world images demonstrate that our SRDN network can achieve competitive results in comparison with the benchmarked and conventional approaches for single-image rain removal.