融合Mamba的多尺度遥感图像实例分割网络

Key Points

• The aim is to improve instance segmentation accuracy in complex remote sensing images with dense target distributions.
• Developed a local-global perception hybrid module for feature extraction.
• Implemented a multiscale deformable convolution to capture target details.
• Utilized Fourier transform for decoupling global frequency components.
• Introduced a hybrid convolution self-attention mechanism to improve segmentation robustness.
• Applied a sliding loss function with exponential moving average to address class imbalance.
• Achieved segmentation accuracies of 65.2% and 68.3% on ISAID and NWPU VHR-10 datasets, respectively.
• Enhanced feature representation and segmentation precision in complex backgrounds.
• Improved the balance between segmentation accuracy and parameter count.