基于GB-YOLO v5su的轻量化小麦麦穗检测方法

Lightweight Wheat Head Detection Method Using GB - YOLO v5su

摘要:

针对复杂农田环境中因植株遮挡及不同生长阶段形态差异导致的小麦麦穗检测精度不足问题，本文提出一种融合轻量设计与多尺度特征增强的检测算法 GB-YOLO v5su (Ghost and BiFPN-enhanced YOLOv5su)。该算法以 YOLO v5su (YOLO v5 small anchor-free) 为基线模型，采用 Ghost 模块重构骨干网络，以构建轻量化 GhostNet 结构，在保持模型表征能力的同时显著降低计算复杂度与参数量;采用加权双向特征金字塔网络 (BiFPN) 替换原 PANet 结构，通过跨尺度连接与可学习的自适应权重机制，增强多尺度特征融合能力;同时新增 P2 检测层构建四尺度特征金字塔，充分利用高分辨率浅层特征提升对小尺度麦穗的检测灵敏度;引入 MPDIoU 损失函数替代 CIoU, 通过最小化预测框与真实框角点距离提升定位精度。在全球小麦检测数据集 GWHD2021 上的实验结果表明，本文算法在参数量仅为 5.38×10?时，平均精度 AP50 达到 92.8%, 实现了精度与效率的良好平衡。系统的消融实验与跨数据集测试进一步验证了各模块的有效性及算法的强鲁棒性。本研究为实现资源受限田间边缘计算场景下的实时精准麦穗检测，提供了一种高效、实用的轻量化部署方案。

Abstract:

Aiming to address the issue of insufficient detection accuracy for wheat heads in complex field environments caused by plant occlusion and growth stage variations, a detection algorithm named GB YOLO v5su, was proposed which integrated lightweight design with multi-scale feature enhancement. Building upon the YOLO v5su model, the algorithm reconstructed the backbone network by incorporating Ghost modules to form a lightweight GhostNet structure, substantially reducing computational complexity and model parameters while maintaining representational capacity. The original PANet was replaced with a weighted bidirectional feature pyramid network (BiFPN), which enhanced multi-scale feature representation through efficient cross-scale connections and a learnable adaptive weighting mechanism. Subsequently, a P2 detection layer was introduced to construct a four-scale feature pyramid, effectively leveraging high-resolution shallow features to improve detection sensitivity for small-scale wheat heads. Furthermore, the MPDIoU loss function was adopted to replace CIoU, which optimized bounding box regression by directly minimizing the corner point distance, thereby improving localization accuracy. Experimental results on the public GWHD2021 dataset demonstrated that the proposed algorithm reduced the model parameter count to 5.38×10? while achieving an average precision (AP50) of 92.8%, striking an effective balance between accuracy and efficiency. Extensive ablation studies and cross-dataset tests further validated the contribution of each module and confirmed the strong robustness of the algorithm. The research result can provide an efficient, practical, and lightweight deployment solution for achieving realtime and accurate wheat head detection in resource-constrained field edge computing scenarios.