ESA-Det: edge and spatial adaptive network for robust object detection in remote sensing imagery

Object detection in optical remote sensing imagery is challenged by the tension between semantic abstraction and the preservation of fine structural cues, particularly for small, elongated, and low-contrast targets immersed in clutter. Deep convolutional hierarchies tend to weaken boundary information during progressive downsampling, which often leads to localization instability and reduced robustness. To address this representational degradation, we propose ESA-Det (Edge and Spatial Adaptive Detector), a lightweight framework that reinforces spatial structure while maintaining high computational efficiency. ESA-Det incorporates two complementary components: (1) an Edge-Prior Infused Attention (EPIA) module that introduces fixed directional gradient cues to stabilize boundary responses within the feature hierarchy, and (2) a Spatial Adaptive Fusion (SAF) block that progressively refines intermediate representations through compact, selectively fused spatial calibration. Extensive experiments on DOTA-v1.5, HRSC2016, and NWPU-VHR-10 demonstrate competitive performance, achieving mAPs of 74.8%, 95.4%, and 92.9% with only 6.6 GFLOPs and 2.89 M parameters. These results suggest that explicit edge priors and adaptive spatial refinement provide a promising direction toward structurally robust remote sensing object detection under complex imaging conditions.