Authors

Abstract

Rapid and accurate detection of millet ears is essential for yield estimation and phenotypic studies. However, traditional detection methods primarily rely on manual observation, which are both subjective and labor-intensive. To address this issue, this study employed Unmanned Aerial Vehicle (UAV) for image data collection of millet ears and proposed the YOLOX-CBAM-EIoU model to facilitate real-time detection, focusing on challenges such as small millet ears size, dense distribution, and severe occlusion in the dataset. Firstly, the Mosaic data augmentation technique was employed to enhance the diversity of the dataset. Subsequently, the CBAM attention mechanism was incorporated between the Neck and Prediction layers of YOLOX, enabling the reallocation of channel weights to enhance the extraction of fine-grained features and deeper semantic information. Additionally, EIoU Loss was utilized as the loss function for bounding box regression to mitigate missed detections in dense scenes. The improved model achieved an average precision (AP) of 90.30%, a 6.44 percentage point increase over the original YOLOX model, significantly enhancing detection performance for densely distributed millet ears. The improved model also demonstrated a Precision of 91.01%, Recall of 89.45%, and F1-score of 90.22, highlighting strong robustness and generalization capabilities. These findings substantiate the efficacy of the YOLOX-CBAM-EIoU model in improving detection performance under dense distribution and occlusion conditions, providing valuable technical reference for further UAV-based analyses of millet ears phenotypes and yield predictions.

Abstract in Chinese

小米穗的快速准确检测对于产量估计和表型研究至关重要。然而，传统的谷穗检测主要依靠人工观察，不仅主观性强且耗时耗力。为此，本研究通过无人机进行谷穗图像数据采集，主要针对数据集中谷穗体积小、分布密集、遮挡严重等问题提出了YOLOX-CBAM-EIoU模型对谷穗进行实时检测。该模型首先在YOLOX的颈部层和预测层之间引入了CBAM注意力模块，通过重新分配不同通道的权重，获得了更浅层的细粒度特征和更深层的语义信息，以提高对谷穗表型的特征提取能力；其次，采用EIOU 函数作为回归损失函数，以改善密集场景下谷穗目标的漏检问题。结果表明，改进后的模型检测平均精度（AP）达到90.30%，与原YOLOX模型相比提高了6.44个百分点，显著提高了密集分布的谷穗目标检测性能。改进后模型的精确率达到91.01%，召回率达到89.45%，F1分数达到90.22，表现出较强的鲁棒性和泛化能力。结果充分证明了YOLOX-CBAM-EIoU模型能显著提高谷穗在密集分布及遮挡条件下的检测效果，为进一步使用无人机分析谷穗表型和产量预测提供了技术参考。