基于计算机视觉的大豆与玉米种子计数方法研究

Research on Soybean and Maize Seeds Counting Method Based on Computer Vision

【目的】作物种子的重量是产量构成的重要因素之一,而传统百粒重/千粒重计算过程耗时、费力,急需一种快速测定作物种子数量、计算重量的方法。【方法】以大豆和玉米为研究对象,首先针对种子计数环境复杂、目标小以及密度大等问题,采用Albumentations库对数据集进行增强处理;然后通过对比YOLOv8的5个子模型,筛选出表现最佳的YOLOv8n模型,在此基础上用Focal-IOU替代CIOU损失函数,得到改进后的模型;最后将改进后的模型与多种经典目标检测模型作对比。【结果】改进后的模型在大豆和玉米种子计数上的平均精度mAP50-95分别达到88.78%和86.89%,比原模型提高了1.29%和0.51%,且性能显著优于YOLOv5、SSD等目标检测模型。此外,改进后的模型在2种作物测试集上的平均绝对百分比误差(MAPE)分别为0.035%和0.045%,每秒帧率分别达到70.17和100.41。【结论】改进后的模型在大豆玉米种子计数上的结果与实际数量差异不显著,实时处理速度快,研究结果可以满足考种中百粒重和千粒重计算对种子的计数需求。

【Objective】The weight of crop seeds is one of the important factors in yield composition,and the traditional calculation process for hundred-grain weight/thousand-grain weight is time-consuming and laborious,urgently requiring a fast method for measuring the number of crop seeds and calculating weight.【Method】This article focuses on soybeans and maize as research objects.Firstly,addressing is⁃sues such as complex seed counting environments,small targets,and high density,the Albumentations library is used to enhance the dataset;Then,by comparing the five sub-models of YOLOv8,the best performing YOLOv8n model was selected.Based on this,Focal-IOU was used to replace the CIOU loss function,resulting in an improved model;Finally,the improved model was compared with various clas⁃sic object detection models.【Result】The results showed that the average accuracy mAP50-95 of the im⁃proved models for soybean and maize seeds datasets reached 88.78%and 86.89%,respectively,which was 1.29%and 0.51%higher than the original model.The performance was significantly better than other object detection models such as YOLOv5 and SSD.In addition,the Mean Absolute Percentage Error(MAPE)of the improved model on the two crop test sets were 0.035%and 0.045%,respectively,and the frame rates per second respectively reached 70.17 and 100.41,respectively.【Conclusion】In conclu⁃sion,the difference between the results of the improved model and the actual quantity in soybean and corn seeds counting is not significant.This model has a fast-real-time processing speed for data.The re⁃search results can meet the needs of seed counting in the calculation of hundred-grain weight/thousandgrain weight in seed testing.