基于深度学习的稻粒在穗计数方法研究  

Study on Counting Method of on Panicle Rice Grains Based on Deep Learning

作　　者：周云成[1] 张羽 刘泽钰 李瑞阳 ZHOU Yuncheng;ZHANG Yu;LIU Zeyu;LI Ruiyang(College of Information and Electronic Engineering,Shenyang Agricultural University,Shenyang 110161,China)

机构地区：[1]沈阳农业大学信息与电气工程学院,沈阳110161

出　　处：《沈阳农业大学学报》2025年第1期82-91,共10页Journal of Shenyang Agricultural University

基　　金：国家重点研发计划项目(2021YFD1500204,2023YFD1501303)。

摘　　要：[目的]稻粒计数是水稻考种的重要环节。针对传统稻穗穗粒人工计数存在着效率低、易出错等问题,研究构建一种稻穗穗粒原位计数模型。原位计数方法可以不破坏稻穗原有拓扑结构,进而进一步应用于其他表型参数获取。[方法]模型以ResNet作为骨干网络,应用图像和范本稻粒之间的特征相关性,预测稻粒概率密度分布,进而通过密度图求和获取稻粒数量。构建稻穗图像数据集,定义稻粒在穗计数模型的损失函数,该函数同时考虑预测密度图与实际稻粒分布的一致性,以及范本标注框的相关约束。用平均绝对误差(Mean Absolute Error,MAE)、均方根误差(Root Mean Squared Error,RMSE)、平均相对误差(Mean Relative Error,MRE)作为衡量模型性能的评估指标。[结果]以ResNet50作为模型骨干网络,模型可取得较理想的精度,其MAE、RMSE、MRE分别为10.937,19.286,13.4%,该方法有着较为准确的计数性能。与YOLOv8-seg相比,本研究模型的MRE下降2.2%,与基于实例分割的SAM(Segment Anything Model)模型相比,本研究模型的MRE降低了12.2%,与T-Rex2模型相比,则降低6.5%。[结论]基于深度学习模型构建,能够自动识别和计数图像中稻粒,提高计数效率,同时相较于其他深度学习模型,本研究模型具有更强的少样本学习能力。本方法可应用于稻粒在穗计数任务,可为稻穗表型参数获取等提供一定技术参考。[Objective]Rice grain counting is a crucial step in rice seed testing.To address the inefficiencies and errors associated with traditional manual counting of rice grains,this study constructed an in-situ counting model for rice grains.In-situ counting method can not destroy the original topological structure of rice panicles,and then it can be further applied to obtain other phenotypic parameters.[Methods]The model employs ResNet as its backbone network and predicts the probability density distribution of rice grains by leveraging the feature correlation between image and rice grain exemplars.Subsequently,the number of rice grains is obtained by summing the density maps.An image dataset of rice panicles was collected,and a loss function tailored for rice grain counting on panicles was defined.This function takes into account both the consistency between the predicted density map and the actual rice grain distribution,as well as the relevant constraints of the exemplar labeling box.The performance of the model was evaluated using Mean Absolute Error(MAE),Root Mean Squared Error(RMSE),and Mean Relative Error(MRE).Experimental[Results]Utilizing ResNet50 as the model's backbone network achieves impressive accuracy,with MAE,RMSE,and MRE values of 10.937,19.286,and 13.4%,respectively.This method exhibits superior counting performance.Compared to YOLOv8-seg,the MRE of the proposed model is reduced by 2.2%.When compared to the SAM(Segment Anything Model)based on instance segmentation,the MRE of the proposed model decreases by 12.2%,and compared to the T-Rex2 model,it is reduced by 6.5%.[Conclusion]This research method is based on the deep learning model,which can automatically identify and count rice grains in the image and improve the counting efficiency.At the same time,compared with other deep learning models,this research model has stronger learning ability with few samples.The method presented in this paper can be effectively applied to the task of rice grain counting in panicles,and the research provides valuable i

关 键 词：稻穗表型 原位计数 水稻 特征提取 表型参数 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] S126[自动化与计算机技术—控制科学与工程]