SIDNet: A single image dedusting network with color cast correction

Abstract

Dust degrades image content and causes image color cast, which negatively impacts on many high-level computer vision tasks. In this paper, we proposed a dedusting network with color cast correction for a single dusty image (SIDNet). The SIDNet contains several dust-aware representation extraction (DustAre) modules with the same structure. Each DustAre module contains two branches. The first branch encodes the input to estimate global veiling-light and local spatial information. The second branch generates a dust-aware map and fuses the global veiling-light, the local spatial information and the dust-aware map to generate the output. To further improve real dusty image dedusting performance, the SIDNet introduces a color cast correction scheme to our neural network. After considering that the average chromaticity values of a dusty image in CIELAB color space are usually larger than those of a clean (dust-free) image, the SIDNet defines a new loss function to better guide the network training. Additionally, we also construct a new synthetic dusty image dataset for network training, which additionally considers the scene depth relationship between real dusty image and dust-free image. Experiments on synthetic and real dusty images show that the SIDNet achieves better dedusting performance compared to state-of-the-art image restoration methods.