基于改进YOLO v5的豆田杂草分布研究

Distribution of Weeds in Soybean Fields Based on Improved YOLO v5

为实现大豆杂草数量和面积的及时精确估算,提出了一种基于改进YOLO v5的大豆田杂草识别方法。以自然场景下的大豆田间杂草为研究对象,利用无人机获取图像数据并进行数据增强;通过引入自适应特征融合机制构建检测模型,结合实测数据建立实测数量面积与估算数量面积线性回归模型,得出农田杂草分布图。对比不同方法对杂草目标提取的结果表明:改进后YOLO v5-ASFF模型优于YOLO v5l、YOLO v5s和YOLO v5x,其F1值为0.903,识别速度为0.268 s/张;实测与估算杂草数量面积线性回归模型相关系数R^(2)=0.9730,拟合度较高。模型误差较低,能够快速、准确地识别大豆杂草数量面积,可为农田范围草情判断提供支撑。

In order to realize the timely and accurate estimation of the number and area of weeds in soybean fields,a method for identifying weeds in soybean fields based on improved YOLO v5 was proposed.Taking weeds in soybean fields in natural scenes as the research object,the image data was acquired and enhanced by UAV.By introducing the adaptive feature fusion mechanism to build a detection model,based on the measured data to establish a linear regression model between the measured quantity and area and the estimated quantity and area,and established the distribution map of farmland weeds.The results of weed target extraction by comparing different methods showed that the improved YOLO v5-ASFF model was better than YOLO v5l,YOLO v5s and YOLO v5x,with F 1 value of 0.903 and recognition speed of 0.268 s/piece.The correlation coefficient R^(2)of the linear regression model between the measured and estimated weed number and area was 0.9730,with a high fitting degree.The method has low error,can quickly and accurately identify the number of soybean seedlings,and can provide support for grass condition judgment in the field.