MSO-Det: A multi-level synergistic optimization framework for robust spike object detection in field scenes

Wang, Yiqun, Zhang, Shuo, Liu, Xiaolong, Ji, Ze ORCID: https://orcid.org/0000-0002-8968-9902 and Jia, Weikuan 2026. MSO-Det: A multi-level synergistic optimization framework for robust spike object detection in field scenes. Computers and Electronics in Agriculture , 112473. 10.1016/j.compag.2026.112473

Full text not available from this repository.

Abstract

Spike object detection is important for precision agriculture, yet existing methods still struggle with density variation, boundary ambiguity, and cross-domain generalization in real field scenes. To address these issues, MSO-Det is proposed as a multi-level synergistic optimization framework, improving spike detection at the feature encoding, multi-scale fusion, and loss optimization levels. Specifically, Adaptive Scene Wave Block (ASWB) replaces encoder self-attention with scene-adaptive frequency-domain wave propagation, enabling the model to adjust its response from sparse to dense scenes. Multi-Scale Interactive Aggregation (MSIA) performs hypergraph-based multi-scale semantic aggregation, capturing long-range cross-level interactions and domain-invariant structure beyond fixed grid neighborhoods. Uncertainty-Guided Distribution Refinement (UGDR) derives uncertainty from predicted side-distance distributions and uses curriculum-guided weighting to improve boundary refinement without introducing inference-time overhead. On the GWHD benchmark, MSO-Det improves AP from 22.2% to 32.2%, [email protected] from 13.1% to 27.8%, and WDA by 9.9 percentage points. It also yields consistent gains across growth stages, density bins, and occlusion levels, and reaches 24.9% AP on the unseen DRPD rice dataset in wheat-to-rice transfer. These results demonstrate that coordinated optimization across representation, fusion, and supervision can substantially improve robustness in field spike detection. The code and models are available at https://github.com/autumn-1010/MSO-Det.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Engineering
Publisher:	Elsevier
ISSN:	0168-1699
Date of Acceptance:	20 September 2026
Last Modified:	12 Oct 2026 11:00
URI:	https://orca.cardiff.ac.uk/id/eprint/190142

Actions (repository staff only)

Edit Item