He Haotian, Ma Xiaodan, Guan Haiou, Wang Feiyi, Shen Panpan
College of Information and Electrical Engineering, Heilongjiang Bayi Agricultural University, Da Qing, China.
Front Plant Sci. 2023 Jan 13;13:1096619. doi: 10.3389/fpls.2022.1096619. eCollection 2022.
As a leaf homologous organ, soybean pods are an essential factor in determining yield and quality of the grain. In this study, a recognition method of soybean pods and estimation of pods weight per plant were proposed based on improved YOLOv5 model. First, the YOLOv5 model was improved by using the coordinate attention (CA) module and the regression loss function of boundary box to detect and accurately count the pod targets on the living plants. Then, the prediction model was established to reliably estimate the yield of the whole soybean plant based on back propagation (BP) neural network with the topological structure of 5-120-1. Finally, compared with the traditional YOLOv5 model, the calculation and parameters of the proposed model were reduced by 17% and 7.6%, respectively. The results showed that the average precision (AP) value of the improved YOLOv5 model reached 91.7% with detection rate of 24.39 frames per millisecond. The mean square error (MSE) of the estimation for single pod weight was 0.00865, and the average coefficients of determination R between predicted and actual weight of a single pod was 0.945. The mean relative error (MRE) of the total weight estimation for all potted soybean plant was 0.122. The proposed method can provide technical support for not only the research and development of the pod's real-time detection system, but also the intelligent breeding and yield estimation.
作为叶片同源器官，大豆荚果是决定籽粒产量和品质的重要因素。本研究基于改进的YOLOv5模型，提出了一种大豆荚果识别方法及单株荚果重量估算方法。首先，利用坐标注意力（CA）模块和边界框回归损失函数对YOLOv5模型进行改进，以检测并准确计数活体植株上的荚果目标。然后，基于拓扑结构为5-120-1的反向传播（BP）神经网络，建立预测模型以可靠估算大豆单株产量。最后，与传统YOLOv5模型相比，所提模型的计算量和参数分别减少了17%和7.6%。结果表明，改进后的YOLOv5模型平均精度（AP）值达到91.7%，检测速率为24.39帧/毫秒。单荚果重量估算的均方误差（MSE）为0.00865，单荚果预测重量与实际重量之间的平均决定系数R为0.945。所有盆栽大豆植株总重量估算的平均相对误差（MRE）为0.122。所提方法不仅可为荚果实时检测系统的研发提供技术支持，还可为智能育种和产量估算提供技术支持。
