A dual branch network combining detail information and color feature for remote sensing image dehazing

Remote sensing images are often affected by haze, resulting in problems such as blurriness, loss of details, and color casts. To effectively remove haze and obtain high-quality remote sensing image, a dual branch network, named DICFNet, that effectively combines detail information and color features is proposed. Specifically, a detail information learning branch is designed firstly, which uses the detail feature residual extraction module (DFREM) to capture the detail features and promote feature learning. Secondly, to learn comprehensive color features, a color feature learning branch is designed. It converts the RGB color space into the Lab color space that is very similar to human visual perception, and then puts the color feature extraction module (CFEM) into use to learn brightness and saturation features. Finally, a learnable fusion module is adopted to obtain the optimal fusion scheme for the previous two branches, enhancing the ability of the model to produce clear remote sensing images. A wealth of experimental evidence indicates that the proposed DICFNet outperforms comparison methods in both visual quality and quantitative evaluation while maintaining a lower memory footprint and requiring fewer computational resources. In addition, detailed ablation experiments demonstrate the effectiveness of the core components of the model.