基于特征重校准与多尺度融合的麦穗计数方法

Wheat Ear Counting Method Based on Feature Recalibration and Multi-scale Fusion

摘要:

在现代化精准农业中,无人机遥感技术凭借高效、低成本等优势,已成为麦穗计数任务的重要手段。然而,无人机影像下麦穗目标尺寸较小,且飞行高度变化导致麦穗尺度不一,同时麦穗在田间分布密集,这些都增加了麦穗计数的难度。针对以上问题,本文提出了一种基于特征重校准与多尺度融合的小麦麦穗计数网络(Featurerecalibration and fusion network,FReCalNet),提升高密度分布与多尺度变化下的麦穗计数精度。具体而言,首先设计了多尺度特征金字塔注意力融合(Feature pyramid attention fusion,FPAttnFusion),加强特征层之间的信息交互,缓解因麦穗尺度变化导致的特征不匹配问题;其次,结合提出的特征重校准与融合模块( Feature recalibration andfusion,FReCal),进一步抑制因麦穗分布密集引起的特征干扰与漏检;最后,构建了多维特征对齐模块(Multi-dimensional feature alignment,MDFA),通过多尺度信息聚合提升模型在不同场景下的鲁棒性和泛化能力。大量实验结果表明,本文方法的平均相对误差(Mean relative error,MRE)、平均绝对误差(Mean absolute error,MAE)和均方根误差(Root mean square Error,RMSE)分别达到12.30%、14.64和17.26,相比性能最优的主流算法TPH-YOLO分别降低了4.5%、4.4%和7.4%;相比常用检测方法FamNet则分别降低了10.0%、10.8% 和11.6%。本研究为高密度麦穗场景下的精准计数提供了有效的技术路径,同时为无人机智能监测与农业精准管理奠定了基础。

Abstract:

In modern precision agriculture, unmanned aerial vehicle (UAV) remote sensing technology has become an essential tool for wheat ear counting due to its high efficiency and low cost. However, UAV imagery presents several challenges: wheat ears appear small in size, their scales vary with flight altitude, and they are densely distributed in the field, all of which increase the difficulty of accurate counting. To address these challenges, a novel wheat ear counting network, feature recalibration and fusion network (FReCalNet) was proposed, which enhanced counting accuracy under conditions of high-density distribution and multi-scale variation. Specifically, a feature pyramid attention fusion (FPAttnFusion) module was designed to strengthen information interaction across feature layers, mitigating feature misalignment caused by scale variations of wheat ears. In addition, a feature recalibration and fusion (FReCal) module was introduced to suppress feature interference and missed detections arising from densely packed wheat ears. Finally, a multi-dimensional feature alignment (MDFA) module was constructed to aggregate multi-scale information and improve the model's robustness and generalization across diverse scenarios. Extensive experimental results demonstrated that the proposed method achieved a mean relative error (MRE) of 12.30%, a mean absolute error (MAE) of 14.64, and a root mean square error (RMSE) of 17.26, representing reductions of 4.5%, 4.4%, and 7.4%, respectively, compared with that of the best-performing baseline method TPH-YOLO. Moreover, compared with the commonly used detection method FamNet, the proposed method achieved reductions of 10.0%, 10.8%, and 11.6%, respectively. The research result can provide an effective technical approach for accurate wheat ear counting in high-density scenarios and lay a solid foundation for intelligent UAV-based monitoring and precision agricultural management.