摘要
木射线作为木材显微构造中的一个重要特征,在木材构造和树种识别的研究中具有重要意义。以针叶材为研究对象,基于计算机视觉技术对木射线特征快速提取方法展开研究,最终构建一套完整的算法,从而使研究人员在向计算机传入针叶材的弦切面图像和比例尺后,即可快速得到木射线细胞数和高度的微观特征。首先对50种针叶材树种的弦切面图像数据集构建U-Net网络模型进行语义分割,对比了在不同放大倍数下木射线的分割效果,发现放大倍数大于10倍时分割效果较好;在此基础上,设计和编写计算分割后图像中的木射线最小覆盖圆算法,可以快速提取木射线的高度,进而得到射线平均高度;再将分割后的木射线图像数据集进行目标检测模型训练,实现射线细胞的目标检测及自动计数功能,并将二值图像阈值分割法和YOLOv3、YOLOv5算法进行比较。结果表明,YOLOv5对射线细胞的检测效果最好。采用本研究方法所提取的木射线高度和细胞数与人工实测结果相比,木射线高度值的偏差不超过4%,细胞计数误差不超过6%。
As an important feature of wood microstructure,wood ray is of great significance in studies of wood structure and tree species identification.The traditional research method is to count the number of ray cells manually.However,due to the large number of ray cells,manual counting is a tedious job and time consuming,as well as easy to miscount.In this study,using softwood as the research object,based on computer vision technology,the computer algorithms and complete package of software for rapid extraction of wood ray features were developed,so that the wood microstructure of the number and heights of wood ray cells could be quickly obtained after inputting the tangential section images and the plotting scales.The specific steps were as follows:firstly,a U-Net network model was constructed for the semantic segmentation of the string section image data set of 50 softwood tree species,and the segmentation effect of wood ray under different magnification was compared.It was found that the segmentation effect was more significant when the magnification was greater than 10 times.On this basis,the incomplete wood ray at the edge of the image was eliminated using the flood-fill method to ensure the integrity of the wood ray,and the minimum co-ver circle algorithm was designed and utilized to calculate the height of the wood ray in the segmentation image,which could quickly extract the height of the wood ray,and then gain the average height of the wood ray.Afterwards,the segmentation wood ray image data set was trained for target detection model to realize target detection and automatic counting functions of ray cells.The binary image threshold segmentation method was compared with the YOLOv3 and YOLOv5 algorithms,and the results showed that YOLOv5 had the best detection outcome on ray cells and the accuracy rate could reach 96.1%.The error between the proposed method and manual statistics was tested.Compared with the measured results,the deviation of ray height and cell count was less than 4%and the cell count error was l
作者
席靖宇
王宇轩
衡利辰
潘彪
骆嘉言
石江涛
王新洲
XI Jingyu;WANG Yuxuan;HENG Lichen;PAN Biao;LUO Jiayan;SHI Jiangtao;WANG Xinzhou(College of Materials Science and Engineering,Nanjing Forestry University,Nanjing 210037,China;Pizhou Markey Supervision and Inspection Center,Pizhou 221300,China)
出处
《林业工程学报》
CSCD
北大核心
2023年第3期132-140,共9页
Journal of Forestry Engineering
基金
江苏省市场监督管理局科技项目(KJ2023048)。
关键词
木材微观构造
木射线
深度学习
语义分割
目标检测
wood microstructure
wood ray
deep learning
semantic segmentation
target detection
