Multi-Resolution Feature Learning with PVT V2 and Parallel Attention Networks in Aerial Imagery

Salient object detection in optical remote sensing images faces significant challenges. These include complex backgrounds, occlusions, and lighting variations. This paper presents GeleNet-X, a novel framework addressing these issues effectively. It combines PVTv2 transformers with advanced attention mechanisms. Key innovations include the Parallel Multiscale Attention Mechanism and Knowledge Transfer Module. These enable superior accuracy in dynamic, real-world scenarios. The proposed system captures both local and global features for improved detection. It leverages multiscale attention to identify objects across varying sizes and orientations. GeleNet-X also integrates edge-aware refinement for sharper object delineation. Experimental results demonstrate state-of-the-art performance across benchmark datasets. The framework is efficient and adaptable for large-scale applications in diverse environments. Applications span agriculture, disaster response, and environmental monitoring.