ACSFANet: Adaptive Cross-Scale Feature Aggregation Network for miniature defect detection in UAV-based distribution network inspection
- 1Hohhot Power Supply Branch of Inner Mongolia Electric Power (Group) Co., Ltd
Abstract
Reliable defect detection of miniature components in power distribution networks remains challenging due to extreme scale variation, progressive loss of fine-grained spatial details in deep feature hierarchies, and structured background interference from complex field environments. This paper proposes ACSFANet, an Adaptive Cross-Scale Feature Aggregation Network that addresses these issues through three synergistic innovations. First, an adaptive cross-scale feature aggregation module employs deformable scale alignment, channel-spatial joint adaptive weighting, and cross-level gated fusion to enable content-conditioned routing of information across pyramid levels, dynamically suppressing or amplifying cross-scale pathways based on target scale characteristics. Second, a detail-preserving detection branch operating at stride-2 resolution retains high-resolution spatial representations critical for sub-16-pixel defect localization. Third, a scale-balanced loss function coupled with dynamic label assignment ensures that gradient signals from tiny, underrepresented defect instances are neither overwhelmed by larger targets nor artificially inflated. Experiments on the DNMD dataset—12,836 UAV-captured images across six defect categories—show that ACSFANet reaches 50.3% mAP and 35.8% APsmall (the COCO definition, i.e. ground-truth boxes with area below 32 × 32 pixels). On mAP the model exceeds the strongest baseline in our comparison, YOLO11-L, by 2.3% points, and on APsmall it exceeds the strongest APsmall baseline, again YOLO11-L, by 5.3 points; against the transformer-based Co-DETR the same two gaps widen to 4.2 and 7.3 points, so the choice of reference baseline visibly changes how large the improvement looks. Ablation studies confirm complementary contributions from each proposed component, and gate activation analysis reveals that the network autonomously learns target-size-dependent cross-scale routing without explicit supervision.
Related
- See More, Detect More: Investigating Colonic Inspection Techniques and Polyp Detection Using Machine Learning
- Q-DOAS: A Proximity Quenching Assay for Real-Time Detection of Early Protein Aggregation Events
- Development of Untargeted Metabolomics-based Detection Method for Naegleria fowleri: A Year-Long Monitoring of the Microbial Ecology in an Operational Drinking Water Distribution System
- Use of COVID-19 Detection Tests in the Mexican Population from 2020-2022: Frequency and Distribution.
- Implementation of Inspection to Controlling Hazards in the Garment
BibTeX
@article{yun2026acsfanet,
title = {{ACSFANet: Adaptive Cross-Scale Feature Aggregation Network for miniature defect detection in UAV-based distribution network inspection}},
author = {Xiaosa Yun and Jianguo Han and Jing Hu and Hongyan Yun and Yaqin Yun and Bateer and Yuanbo Wen},
journal = {Scientific Reports},
year = {2026},
month = aug,
doi = {10.1038/s41598-026-62985-3},
url = {https://doi.org/10.1038/s41598-026-62985-3},
note = {CC BY},
}
RIS
TY - JOUR
TI - ACSFANet: Adaptive Cross-Scale Feature Aggregation Network for miniature defect detection in UAV-based distribution network inspection
AU - Xiaosa Yun
AU - Jianguo Han
AU - Jing Hu
AU - Hongyan Yun
AU - Yaqin Yun
AU - Bateer
AU - Yuanbo Wen
PY - 2026
DA - 2026/08/01
JO - Scientific Reports
DO - 10.1038/s41598-026-62985-3
UR - https://doi.org/10.1038/s41598-026-62985-3
AB - Reliable defect detection of miniature components in power distribution networks remains challenging due to extreme scale variation, progressive loss of fine-grained spatial details in deep feature hierarchies, and structured background interference from complex field environments. This paper proposes ACSFANet, an Adaptive Cross-Scale Feature Aggregation Network that addresses these issues through three synergistic innovations. First, an adaptive cross-scale feature aggregation module employs deformable scale alignment, channel-spatial joint adaptive weighting, and cross-level gated fusion to enable content-conditioned routing of information across pyramid levels, dynamically suppressing or amplifying cross-scale pathways based on target scale characteristics. Second, a detail-preserving detection branch operating at stride-2 resolution retains high-resolution spatial representations critical for sub-16-pixel defect localization. Third, a scale-balanced loss function coupled with dynamic label assignment ensures that gradient signals from tiny, underrepresented defect instances are neither overwhelmed by larger targets nor artificially inflated. Experiments on the DNMD dataset—12,836 UAV-captured images across six defect categories—show that ACSFANet reaches 50.3% mAP and 35.8% APsmall (the COCO definition, i.e. ground-truth boxes with area below 32 × 32 pixels). On mAP the model exceeds the strongest baseline in our comparison, YOLO11-L, by 2.3% points, and on APsmall it exceeds the strongest APsmall baseline, again YOLO11-L, by 5.3 points; against the transformer-based Co-DETR the same two gaps widen to 4.2 and 7.3 points, so the choice of reference baseline visibly changes how large the improvement looks. Ablation studies confirm complementary contributions from each proposed component, and gate activation analysis reveals that the network autonomously learns target-size-dependent cross-scale routing without explicit supervision.
SN - 2045-2322
ER -