Scholarlink

<p>To address the detection challenges posed by tiny wheat-ear targets, highly similar wheat canopy backgrounds, dense occlusion, and scale variation in low-altitude UAV images, this paper proposes MBCD-YOLO, a lightweight detection algorithm based on YOLO11n. First, a Micro-Feature Preservation Module (MFPM) uses Space-to-Depth rearrangement to preserve fine-grained spatial information during downsampling. Second, a Boundary-Semantic Synergy Block (BSSB) fuses fixed Sobel-gradient priors with learned semantic features to improve boundary discrimination under homogeneous backgrounds. Third, a Context-Aware Reconstruction Module (CARM) uses efficient up-convolution composed of 2x upsampling, 3 × 3 depthwise convolution, channel shuffle, and 1 × 1 pointwise fusion to reduce feature aliasing and spatial misalignment in multi-scale fusion. Finally, a Distribution-Guided Quality Head (DGQH) introduces localization-quality estimation derived from regression distributions to improve dense-object ranking during NMS. Experiments on the Global Wheat Head Detection Dataset (GWHD) show that MBCD-YOLO achieves 50.6% mAP@50–95 and 81.7% recall while maintaining 89 FPS. In the counting-oriented evaluation, MBCD-YOLO achieves an MAE of 4.92, an RMSE of 7.18, and a counting accuracy of 88.33%, demonstrating its effectiveness for dense wheat-ear detection and counting.</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

Tiny target detection algorithm in field wheat canopies based on MBCD-YOLO

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月19日

查看全文
  • Zhangping Li,
  • Zehan Wei

摘要

To address the detection challenges posed by tiny wheat-ear targets, highly similar wheat canopy backgrounds, dense occlusion, and scale variation in low-altitude UAV images, this paper proposes MBCD-YOLO, a lightweight detection algorithm based on YOLO11n. First, a Micro-Feature Preservation Module (MFPM) uses Space-to-Depth rearrangement to preserve fine-grained spatial information during downsampling. Second, a Boundary-Semantic Synergy Block (BSSB) fuses fixed Sobel-gradient priors with learned semantic features to improve boundary discrimination under homogeneous backgrounds. Third, a Context-Aware Reconstruction Module (CARM) uses efficient up-convolution composed of 2x upsampling, 3 × 3 depthwise convolution, channel shuffle, and 1 × 1 pointwise fusion to reduce feature aliasing and spatial misalignment in multi-scale fusion. Finally, a Distribution-Guided Quality Head (DGQH) introduces localization-quality estimation derived from regression distributions to improve dense-object ranking during NMS. Experiments on the Global Wheat Head Detection Dataset (GWHD) show that MBCD-YOLO achieves 50.6% mAP@50–95 and 81.7% recall while maintaining 89 FPS. In the counting-oriented evaluation, MBCD-YOLO achieves an MAE of 4.92, an RMSE of 7.18, and a counting accuracy of 88.33%, demonstrating its effectiveness for dense wheat-ear detection and counting.