春玉米根系原位成像图片自动识别研究

微根窗成像技术推动了植物根系表型研究，但是根系长度和直径仍然需要人眼识别再绘制轨迹，消耗大量的人力和时间。为了解决这一难题，本研究将U-Net语义分割技术引入到植物根系图像识别中，研发了基于机器学习的iRoot-V02根系自动识别软件。采用iRoot-V02软件识别微根窗法获得的植物根系成像图片的根长、直径、投影面积、根尖数等参数。结果表明: iRoot-V02软件批处理600 dpi图片的速度为每张26.6 S; 获取根系的骨架信息、总根长与人眼识别结果基本一致; 按直径每0.1 mm为一级，获得不同直径的根长，与人眼识别结果的根长决定系数大于0.76;精确捕捉到根系生长旺盛期不同直径根长的变化; 分析了300 dpi和600 dpi两种分辨率根系图片的参数，两种分辨率结果具有高相关性，因此可建立低分辨率根系参数和高分辨率根系参数之间的关系方程，采用低分辨率拍摄根系图像，通过方程转化成更准确的根系参数，减轻工作量。用iRoot-V02软件的根系生长信息近似于人眼识别，相比人眼识别在大批量根系图像智能识别、自动化、快速目标检测方面具有巨大优势。

Study on automatic recognition of root system image of spring maize in situ

Minirhizotron technology promotes the study of root phenotype, but the root length and diameter still require human eye recognition to draw the track, which consumes a lot of manpower and time.This study introduced U-Net semantic segmentation technology into plant root image recognition, and developed iRoot-V02 software based on machine learning to solve this problem.The iRoot-V02 software was used to identify the root length, diameter, projection area, and root tip number from plant root imaging images obtained by Minirhizotron technology.The results show that the average speed of iRoot-V02 software for processing 600 dpi images in batches is 26.6 seconds per picture.The skeleton information and total length of roots are essentially consistent with the human eye recognition results.According to the diameter of each 0.1 mm as a level, the correlation coefficient between the obtained root length of different diameters and the eye recognition results is larger than 0.76, reflecting that the software accurately captures the changes of root lengths of different diameters in the vigorous growth period of maize.In addition, the analysis of the parameters of root images with 300 dpi and 600 dpi resolutions shows that their results are highly correlated.Therefore, the relationship equation between low-and high-resolution root parameters can be established to confirm more accurate root parameters with the low-resolution root image and to reduce the workload.Root growth information obtained by using iRoot-V02 software is similar to that of human eye recognition.On a whole, compared with human eye recognition, iRoot-V02 has great advantages in mass root image intelligent recognition, automation and fast target detection.