A Two-Stage Fusion Remote Sensing Image Dehazing Network Based on Multi-Scale Feature and Hybrid Attention

Mengjun Miao,Heming Huang,Feipeng Da,Dongke Song,Yonghong Fan,Miao Zhang
DOI: https://doi.org/10.1007/s11760-024-03160-x
IF: 1.583
2024-01-01
Signal Image and Video Processing
Abstract:Remote sensing images acquired under bad weather conditions often suffer from serious degradation such as color distortion, blur and low contrast, which seriously affects their application in vision tasks. To this end, a two-stage fusion dehazing network, termed TSFDNet, is proposed to remove the haze in remote sensing images effectively. In the first stage, the preliminary dehazing sub-network is designed to remove haze, which employs a multi-scale feature extraction block to aware haze density features to enhance the dehazing effect. In the second stage, the refined dehazing and detail compensation sub-network is designed to refine dehazing and compensate for image details by utilizing edge information and pixel-channel hybrid attention residual modules. Finally, the potentially beneficial features of the two stages are fused to improve the model performance. Experiments on multiple datasets have shown that the proposed model performs better in quantitative and qualitative evaluations than the compared methods. Furthermore, the effectiveness of key components of the model has been verified by ablation studies.
What problem does this paper attempt to address?