Answers are grounded in this paper and cite their sources.
Demonstrates improved wheat head detection in agriculture, indicating potential for real-world application.
Wheat head detection is a cornerstone task for yield estimation and crop counting in precision agriculture. Nevertheless, many challenges including dense distribution, complex backgrounds and varying lighting must be tackled to guarantee detection accuracy. This study enhances wheat head detection performance using a novel coupled data-model optimization approach. First, we replace horizontal bounding boxes (HBBs) of the Global Wheat Head Detection (GWHD) dataset with oriented bounding boxes (OBBs) generated using the Segment Anything Model (SAM), creating the Comprehensive Rotated Wheat Head Detection (C-RWHD) dataset, which offers reduced background complexity and overlapping impacts by emphasizing target information. Secondly, YOLOv11 and YOLOv12 architectures are tailored for wheat detection by removing specific layers to reduce resource utilization. Third, we introduce the Tunisian Wheat Head Detection (TWHD) dataset, self-compiled and specifically assembled for robustness assessment. Results show that compared to its OBB baseline, the proposed YOLOv12s-OBB-SM reduces parameters by 28.4% and Giga Floating-Point Operations (GFLOPs) by 2.2, surpasses it in recall, maintains consistently high performance with 96.3% mean Average Precision at Intersection over Union 0.5 (mAP@0.5), offers 28.6% higher frames per second (FPS), while demonstrating strong generalization on TWHD, and providing strong counting efficiency across a complete non-selective test set of 1382 images with R 2 =0.874 and Mean Absolute Error (MAE)=7.2. Additionally, YOLOv11n-OBB-SM outperforms its OBB baseline in counting, achieving R 2 =0.83 and MAE=8.4 with 29.63% fewer parameters and 11.94% fewer GFLOPs. Overall, the proposed architecturally tailored models meet real-time inference and computational frugalness requirements of agricultural edge devices, demonstrating strong practical potential for field deployment.
No takes yet. Share an insight, caveat, or question.
Khazmi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: