Bridging global and local contexts: a multistage approach for high-resolution remote sensing image segmentation

Semantic segmentation of high-resolution remote sensing images constitutes a critical research area in computer vision. This task presents significant challenges due to the intricate structures in high-resolution images, which increase segmentation complexity, and the large image sizes, which necessitate multiple downsampling operations during model inference, leading to substantial information loss and degradation of segmentation accuracy. To address these challenges, this paper introduces a multi-stage feature enhancement and fusion framework designed specifically for high-resolution remote sensing image segmentation. The proposed framework integrates a multi-stage refinement segmentation model, wherein different stages progressively capture and optimize features at various abstraction levels, thereby enhancing adaptability to diverse segmentation scenarios. Furthermore, a global/local feature enhancement module is incorporated to reinforce both global and local features while mitigating information loss incurred during local feature extraction. Additionally, an adaptive fusion module dynamically assesses the relative importance of different regions, facilitating adaptive feature fusion across both channel and spatial dimensions. Extensive experiments conducted on public benchmark datasets, including DeepGlobe and Inria Aerial, demonstrate that the proposed approach achieves superior segmentation performance across multiple evaluation metrics, outperforming existing state-of-the-art methods for high-resolution remote sensing image segmentation.