Dual‐Domain Fusion Attention Network for Remote Sensing Image Dehazing

ABSTRACT In remote sensing image processing, haze significantly reduces image clarity, thereby severely impacting subsequent image processing and analysis tasks. Convolutional Neural Network (CNN) based dehazing models can effectively remove haze from remote sensing images, thereby enhancing their clarity. However, existing dehazing methods based on CNN have limitations in handling global dependencies. When training deep networks, gradient signals may gradually diminish or even vanish. To address the aforementioned challenges, this paper proposes a novel Dual‐domain Fusion Attention Network (DDFAN) based on the concept of frequency‐spatial dual guidance, leveraging the strengths of Transformers in capturing global relationships. DDFAN incorporates a new Spatial and Channel Fusion Attention (SCFA), complemented by a Multi‐Dconv Head Transposed Attention Expanded by Taylor Formula (T‐MDTA). This combined approach enables DDFAN to recover finer details from hazy images. Experimental results on the remote sensing Haze1k dataset demonstrate that the proposed method achieves significant improvements in dehazing performance metrics, with an average increase of 1.2888 dB in Peak Signal‐to‐Noise Ratio (PSNR) and 0.0139 in Structural Similarity Index Measure (SSIM), representing an improvement of approximately 1.5% over the baseline and outperforming existing methods. Our code is released publicly at https://github.com/gouyihao/DDFAN .