Remote sensing image semantic segmentation integrating feature complementarity and linear attention

The classical Feature Pyramid Network often results in the neglect of detail information due to information downsampling. Furthermore, the context dependency of the input area is homogeneous, which limits the accuracy of image segmentation. To solve the above issues, this article proposes a novel Feature Pyramid Network by incorporating feature complementarity and a linear attention mechanism. Reweighing single-scale features in grid fusion, the network utilizes cross-scale complementary knowledge to decrease the neglect of local details in the image. Additionally, a linear attention mechanism with a differentiable linear unit kernel function is leveraged to enhance the long-range pixel associations across the global image scope. This mechanism adaptively allocates attention weights with a lightweight structure, especially emphasizing critical information such as object boundaries for the enhancement of the proposed model's accuracy and robustness. After testing, the overall accuracy of segmentation is up to 92.7% and 92.9%, respectively, which proves that the proposed method can significantly improve the segmentation performance.