Single-image raindrop removal using concurrent channel-spatial attention and long-short skip connections

Abstract

Image raindrop removal refers to removing raindrops from the images taken through glass under rainy weather. It is a challenging problem due to the large variations of raindrop appearances as well as the nonlinear visual distortions caused by raindrops. This problem becomes much more difficult when only a single image is provided. This paper aims at tackling the single-image raindrop removal problem by leveraging the power of deep learning. We proposed an end-to-end approach using the deep neural network. To address the challenges of raindrop removal, we introduce a concurrent channel and spatial attention mechanism implemented by Squeeze & Excitation into the network. The channel attention mechanism allows individually selecting useful features in the network for handling the raindrops of different appearances and restoring different image patterns. The spatial attention enables different treatments to the image regions with different degrees of distortion and with different local image patterns. In addition, long-short skip connections are added for utilizing the intermediate features of the network for better restoration. The experimental results show that the proposed approach outperformed the existing ones.