LMAPNet: a lightweight multi-scale attention network for real-time maize disease and pest detection on edge devices
Introduction Accurate detection of maize diseases and pests is essential for targeted crop protection, but field performance is limited by scale variation, weak lesion boundaries, occlusion and background interference. Methods This study proposes the Lightweight Multi-Scale Attention and Path-Aggregation Network (LMAPNet), a compact detector based on YOLO26n. The Adaptive Multi-Scale Initialization (AMSI) module combines parallel depthwise receptive fields with branch-wise channel recalibration to preserve features across target scales. The Multi-Scale Attention Path-Aggregation Network (MSAPAN) uses attention-guided bidirectional fusion to integrate high-level semantics with low-level spatial details. Inverted Residual Mobile Block (iRMB)-based refinement further limits irrelevant background responses. Slide -GIoU Loss (SG-Loss) applies sample-quality-aware classification weighting and provides localization signals for non-overlapping bounding boxes without increasing inference-time computation. Results On the held-out test set, LMAPNet improved precision and [email protected] by 3.1 and 2.5 percentage points, respectively, while reducing the parameter count by 27.1% relative to YOLO26n. The complete system was deployed on an NVIDIA Jetson Orin Nano with an IMX219 camera, demonstrating the feasibility of on-device field monitoring under the tested conditions. Discussion LMAPNet provides a compact solution for real-time maize disease and pest detection on resource-constrained agricultural devices.
Authors
- Ke Mei (ORCID: https://orcid.org/0000-0002-2469-0523)
- Wendong Xiao (ORCID: https://orcid.org/0000-0002-9398-0794)
- Zijian Wang
- Hong Jiang
- Ze Zhang
Institutions
- Shihezi University (CN)
- Changji Vocational and Technical College (CN)
- Xinjiang University (CN)
Publication Details
- Journal
- Frontiers in Plant Science
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fpls.2026.2010865
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00