The growing scope of image degradation encountered in remote sensing detection has sparked significant interest in the application of deep learning methods. To address the challenges posed by high-frequency signal loss and structural distortion in reconstructed remote sensing images, a transformer-based skip network with frequency attention (named FATSNet) is proposed to improve the model's ability. The model comprises two key blocks: encoder and decoder blocks, which incorporate two novel frequency attention modules. In the encoder block, a discrete wavelet transform (DWT) module is designed to implement the representation of detailed abstract content features. The decoder block utilizes a frequency-aware transformer (FAT) module based on fast Fourier transform (FFT) to enhance the clarity of the image structures and boundaries. To improve the robustness of the network during feature fusion, a self-adapting weighting module (SAWM) is proposed by dynamically adjusting the weights of feature channels to make more effective use of feature information. The experimental results with public datasets demonstrate the superiority of the FATSNet model over other leading methods. Additionally, a case study of the detection of Lake Buridun was used to verify the practicality of the proposed model.
Huo et al. (Thu,) studied this question.