Human-in-the-loop collaborative enhancement for rotation-aware small object detection in aerial images

Abstract Small object detection in aerial images remains challenging due to low resolution, arbitrary orientation, complex background, and extreme sparsity of targets. Existing deep learning methods struggle with weak feature discrimination and insufficient context modeling for tiny targets, and they largely treat human knowledge as external to the detection pipeline. This paper presents a human-in-the-loop (HITL) collaborative detection framework that combines super-resolution enhancement, rotation-aware detection, and dual human-guided feature augmentation. Concretely, we introduce an enhancement scheme consisting of: (1) a super-resolution-based rotation-aware detection module that improves the perceptual quality and orientation robustness of tiny objects, (2) an HITL categorical feature enhancement module that sharpens target–background discrimination, and (3) an HITL regional feature enhancement module that selectively boosts potential meaningful locations with low additional cost. Compared with FPN-attention-based approaches, our method keeps the detector architecture simple and detector-agnostic, avoiding complex feature-pyramid attention modules and working as a modular wrapper around standard detectors. In extensive UAV surveillance experiments on a tailored and mixed 5-class dataset from DOTA, HRSC2016, and NWPU VHR-10, our method achieves an 85.0% mAP@50, showing competitive performance against representative baselines under the same evaluation protocol. The proposed framework effectively improves detection mAP and interpretability by combining human prior knowledge with deep learning, showing practical value for operator-assisted UAV analysis scenarios. To facilitate reproducibility and further research, the full source code of the proposed algorithm has been released on GitHub and archived on Zenodo ( https://doi.org/10.5281/zenodo.20490958 ).