基于RAPESEED-YOLO的散铺油菜角果籽粒数检测方法研究

Detection Algorithm of Scattered Rapeseed Pods Based on RAPESEED-YOLO

摘要:

油菜角果平均籽粒数是评价产油量的核心指标。传统人工检测方法存在破坏性大、主观误差大、效率低等缺陷。为此,提出一种基于姿态信息的角果内含籽粒数统计方法,实现散铺油菜角果籽粒的自动化无损检测。引入RAPESEED-YOLO 网络,加入轻量化VanillaNet 主干网络、P2小目标检测头及多尺度膨胀注意力机制(MSDA),增强小目标特征提取能力;设计动态检测头(Dynamic Head)与动态采样器(DySample),结合C3k2_MDConv2 模块优 化多尺度特征融合;提出Shape-IoU 损失函数,提升小目标边界框回归精度,降低模型体积,提升运算速度,便于部署。实验结果表明,RAPESEED-YOLO网络精度为98. 6% 、召回率为98. 7% 、平均精度均值为98. 9% ,本统计方法在15组生产实验中平均精度为95. 6% ,误差小于10% ,理想状态下可达到3 株/ h的高通量。展现出在农业场景下高鲁棒性与低资源消耗特性,为油菜籽粒无损检测提供了精准、高效、轻量化的实用技术路径,对推动油菜测产具有重要应用价值。

Abstract:

Rapeseed is China’ s dominant oil crop and the mean seed number per silique is the key predictor of yield; however, manual dissection is destructive, subjective and labor-intensive. A pose- aware, non-destructive counting pipeline that integrated digital imaging with a customized YOLO detector named RAPESEED YOLO was proposed. The network employed a lightweight VanillaNet backbone, a P2 micro-object head and multi-scale dilated attention (MSDA) to enhance tiny-seed features, while a Dynamic Head with DySample up-sampler and a C3k2 _ MDConv2 fusion module refined multi-scale context. Training was regularized by Shape-IoU loss to improve bounding-box regression for ellipsoidal seeds. Evaluated on 15 field plots ( three cultivars, two densities), RAPESEED-YOLO achieved 98. 6% precision, 98. 7% recall and 98. 9% mAP; the overall pipeline delivered 95. 6% counting accuracy with less than 10% relative error across 6 800 siliques. Operating on a 15 W edge device, the system processed three plants per hour, demonstrating robustness to uneven illumination, occlusion and cultivar variation. Pose-specific multiplication factors ( upright 1 × , semi-lateral 2 × strong-side, full- lateral 2 × single-side) convert detected seeds into total silique counts without physical contact. With only 2. 3 × 106 parameters and 5. 1 × 109 FLOPs, the method offered a lightweight, accurate and deployable solution for high-throughput rapeseed yield estimation.