Abstract: Haze interference can cause blurred structures and loss of details in remote sensing images, severely compromising the accuracy of downstream visual tasks. To address this challenge, this paper proposes a heterogeneous enhancement remote sensing image dehazing network that improves feature restoration from two perspectives: spatial structure modeling and frequency information integration. Specifically, a multi-scale asymmetric attention transformer module is designed, incorporating a direction-aware mechanism to enhance the modeling of blurred edges and texture details. In parallel, a wavelet-based adaptive high-low frequency enhancement module is constructed, utilizing Haar wavelet decomposition to separate frequency-domain information, where high-frequency and low-frequency submodules are employed to reinforce edge contours and structural representations, respectively. These two modules are embedded in the feature extraction and feature fusion stages, collaboratively addressing the limitations of traditional methods in directional modeling and high-frequency feature preservation. With low computational overhead, the proposed method achieves average PSNR/SSIM scores of 24.993 6/0.909 9 on the HAZE1K dataset and 33.180 2/0.894 2 on the RICE dataset, demonstrating significant advantages in detail restoration.

Key words: remote sensing image dehazing, transformer, asymmetric attention, high-low frequency feature enhancement, wavelet transform, direction-aware modeling, deep learning