Scholarlink

<p>Wheat spike detection is essential for yield estimation and field management, yet real-world field images exhibit large-scale variation, extreme density, heavy occlusion, and strong background textures, which often cause false and missed detections. To address these challenges, we build on the RT-DETR-R18 baseline and propose a structurally enhanced detector tailored for dense small-object wheat-spike detection. Our method reduces omissions and misrecognitions by introducing a self-devised Small Object Enhancement Pyramid (SOEP) that strengthens high-resolution representations and multi-scale feature fusion, including SPDConv-based P2 enhancement, CSP-OmniKernel integration, and replacing the original concatenation-based fusion with a Modulation Fusion Module (MFM); and by strengthening the detection head to better localize and detect densely packed spikes. Ablation studies show that each component yields consistent gains, and the final model achieves a precision of 0.9211 and an mAP<sub>50</sub> of 0.9208 on GWHD2021, demonstrating the effectiveness and superiority of the proposed method for wheat spike detection in complex field scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Improved RT-DETR-Based Algorithm for Small Wheat Spike Detection

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Xuyang Wang,
  • Zhonglin Ji

摘要

Wheat spike detection is essential for yield estimation and field management, yet real-world field images exhibit large-scale variation, extreme density, heavy occlusion, and strong background textures, which often cause false and missed detections. To address these challenges, we build on the RT-DETR-R18 baseline and propose a structurally enhanced detector tailored for dense small-object wheat-spike detection. Our method reduces omissions and misrecognitions by introducing a self-devised Small Object Enhancement Pyramid (SOEP) that strengthens high-resolution representations and multi-scale feature fusion, including SPDConv-based P2 enhancement, CSP-OmniKernel integration, and replacing the original concatenation-based fusion with a Modulation Fusion Module (MFM); and by strengthening the detection head to better localize and detect densely packed spikes. Ablation studies show that each component yields consistent gains, and the final model achieves a precision of 0.9211 and an mAP50 of 0.9208 on GWHD2021, demonstrating the effectiveness and superiority of the proposed method for wheat spike detection in complex field scenarios.