ADMSF-YOLO: structural refinement and prediction-scale reallocation for UAV small-object detection
Small-object detection in unmanned aerial vehicle (UAV) imagery remains challenging because distant targets often occupy only a few pixels after image resizing and hierarchical feature downsampling. As a result, fine contours and local textures are easily weakened, while the standard P 3 – P 4 – P 5 prediction hierarchy assigns part of the detection capacity to coarse feature maps that are less suitable for dense tiny targets. The present study therefore focuses primarily on improving detection accuracy for such targets while explicitly evaluating the associated computational cost. To address this mismatch, this paper proposes ADMSF-YOLO, a YOLO11s-based detector that integrates structural refinement, learnable normalized cross-scale fusion, pre-upsampling feature modulation, and scale-guided prediction-head reallocation. In the backbone, DSR-C3k2 applies context-detail residual compensation before feature aggregation and structural recalibration afterward. In the neck, AMSF uses module-specific normalized scalar weights learned during training to combine aligned feature branches, while LRFM-Up modulates low-resolution semantic features before parameter-free nearest-neighbor upsampling. The prediction hierarchy is further reallocated from P 3 – P 4 – P 5 to P 2 – P 3 – P 4 , shifting prediction toward higher-resolution feature maps without increasing the number of detection heads. Experiments on VisDrone2019-DET and AI-TOD show that ADMSF-YOLO improves mAP 50 over YOLO11s by 7.5 and 9.9 percentage points and mAP 50:95 by 5.2 and 5.3 percentage points, respectively. Under the two-class TinyPerson setting used in this study, mAP 50 increases from 0.154 to 0.195 and mAP 50:95 from 0.054 to 0.069. Scale-specific evaluation on VisDrone2019-DET further shows that the P 2 – P 3 – P 4 hierarchy yields higher AP S and AP M than the four-head alternative, while AP L decreases slightly. Relative to YOLO11s, these gains are accompanied by an increase in GFLOPs from 21.5 to 34.4 and a decrease in the measured network forward throughput with float32 inputs from 192 to 138 FPS under the reported protocol.
Authors
- Wei He
- Yu Zhao
- Zhijie Zhou
- Yong Deng
- Guohui Zhou
Institutions
- Harbin Normal University (CN)
- University of Electronic Science and Technology of China (CN)
Publication Details
- Journal
- Journal of King Saud University - Computer and Information Sciences
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1007/s44443-026-01395-x
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00