Dynamic Spatial Perception Network for Slender Target Detection in Aerial Image

ABSTRACT

Rotating object detectors are widely used in remote sensing image analysis, but most existing methods rely on numerous preset rotated anchors to cover targets, neglecting the adverse impacts of high‐aspect‐ratio (HAR) target shape variations on network training. Two key issues arise: (1) Scarcity of high‐quality positive samples for slender targets (critical positional information included), causing insufficient fine‐grained localisation supervision; (2) Abrupt gradient fluctuations during regression, leading to training instability. To address these, we propose an aspect‐ratio‐adaptive sample assignment strategy and a stable HAR bounding box regression training strategy. Specifically, we design the Shape Adaptive Label Assignment strategy with an intersection over union (IoU)‐based weight function to prioritise positive samples matching HAR targets. Secondly, the Gradient Equalisation Regression Loss function is proposed to mitigate HAR target regression gradient instability and improve convergence. Experiments on DOTA and HRSC‐2016 show the proposed method attains 81.2% and 91.8% mAP, respectively, outperforming state‐of‐the‐art aerial detectors.