摘要: 目的为降低原木检尺作业中人为因素对检尺结果的影响，提升工作效率，提出一种基于掩膜区域实例分割模型和边缘拟合算法的原木径检尺方法、�/sec> 方法使用单目手机作为采集设备，针�?种不同尺径等级的桉树原木和矩形标尺作为研究对象。首先在不同距离下采集图像制作数据集，以8�?�?比例划分训练集、验证集和测试集，建立原木端面识别实验数据集。其次利用实例分割模型提取端面部分生成掩膜，使用边缘拟合算法求得矩形标尺和原木端面像素长度，结合标尺实际大小求得原木端面实际尺径。比较算法测量误差及不同国家标准下材积计算误差，评估该方法的准确性、�/sec> 结果本实例分割模型能够准确地实现原木端面掩膜分割，达�?9.89%的精准率�?9.41%的召回率，F1分数与均值平均精度相较one-stage算法有明显提升。通过最小二乘边缘拟合算法拟合端面为椭圆，求得椭圆短径作为原木尺径，对比真值，平均百分比误差约为−2.00%，较真实值偏小。对比不同尺径等级原木误差，100%小尺径原木�?8%中尺径原木和95%大尺径原木的计算值误差范围为�?% ~ 5%。对比不同距离下采集的原木端面图像，�?0 ~ 100 cm以内采集图像效果最佳，平均相对误差不超过−2.22%，距离大�?00 cm时误差逐步提升。对比不同国家原木材积计算标准，根据我国标准得出材积误差为−4.5%，与美国、俄罗斯和日本的标准相比较，误差更小、�/sec> 结论相较于人工检尺径，本研究提出的测量方法工作效率更高，人为因素影响更小，能够较为准确地测得原木尺径，可以达到代替人工原木检尺径作业的目标、�/sec>

关键诌�
• 原木检尹�/a> /
• 单目视觉/
• 实例分割/
• 边缘拟合算法

Abstract: ObjectiveIn order to reduce the influence of human factors on the log sizing results and improve the work efficiency, a log sizing method based on mask region instance segmentation model and edge fitting algorithm was proposed. MethodThe method used monocular camera as the acquisition equipment, and the end face images of eucalyptus with three different diameter classes and rectangular scaleplate were taken as the research objects. Firstly, the image was collected at different distances to make a data set, and the training set, verification set and test set were divided at the ratio of 8�?�?, and the log end face recognition experiment database was established. Secondly, the instance segmentation model was used to extract the end face part to generate the mask, the edge fitting algorithm was used to obtain the pixel size of the rectangular scaleplate and the log end face, combined the actual size of the scaleplate to obtain the actual size of the log end face. Finally, the error of fitting size of the algorithm and the error of different national standard volume calculation formulas were compared. ResultIt was found that the example segmentation model in this paper can achieve more accurate mask segmentation of log end face, and the accuracy and recall rate were 99.89% and 99.41%, respectively. Compared with the one-stage algorithm, F1 scores and mean average precision were significantly improved. By fitting the end face to an ellipse using a least square edge fitting algorithm, the short diameter of the ellipse was obtained as the log diameter. Compared with the real value, the average percentage error was about �?.00%, which was slightly smaller than the real value. Comparing the error of logs with different sizes, the measurement error of 100% small size logs, 98% medium size logs and 95% large size logs was �?%�?%. By comparing the log end face images collected at different distances, it was found that the best effect was to collect images within 50�?00 cm, the average relative error didn’t exceed �?.22%, and the error gradually increased when the distance was greater than 100 cm. By comparing the calculation standards of log volume in different countries, the volume error was �?.5% according to the calculation formula of wood volume in China, which was the lowest among the four national standards of China, America, Russia and Japan. ConclusionCompared with manual measuring, the measuring method proposed in this paper is more efficient, less affected by human factors, and can more accurately measure the log size, and achieve the goal of replacing manual log sizing operations.

Key words:
• log scale/
• monocular vision/
• instance segmentation/
• edge fitting algorithm

�?nbsp; 1原木端面标注效果

Figure 1.Log end face labeling effect

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 2实例分割模型结构

Figure 2.Structure of instance segmentation model

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 3网络训练损失变化曲线

Figure 3.Network training loss changing curve

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 4识别结果示例

Figure 4.Example of recognition results

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 5拟合结果示例

Figure 5.Example of fitting results

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 6实例分割模型计算值与测量值的检尺径误差分析

N为原木总数＋�i>e为平均百分比误差，SD为标准差、�i>Nis the total number of logs,eis mean percentage error, and SD is the standard deviation.

Figure 6.Error analysis of caliper diameter between calculated and measured values of instance segmentation model

下载: 全尺寸图牆�/a> 幻灯牆�/a>

�?nbsp; 1不同模型整体性能对比

Table 1.Comparison of overall performance of different models

模型 Model	精准玆�br/>Accuracy/%	召回玆�br/>Recall rate/%	F1分数 F1 score	均值平均精�?br/>Mean average precision	平均训练时间 Average training time/min
YOLOv3-tiny	78.32	72.46	0.841	0.877	4
YOLOv4-tiny	93.21	89.56	0.945	0.964	3
SSD	82.47	80.30	0.890	0.923	4
本模� This model	99.89	99.41	0.989	0.978	52

下载: 导出CSV

�?nbsp; 2不同距离下平均尺径误�?/p>

Table 2.Average caliper error at different distances

距离 Distance/cm	平均绝对误差 Mean absolute error/mm	平均百分比误�?br/>Mean percentage error/%
50	�?.62	�?.03
75	�?.79	�?.22
100	�?.35	�?.94
125	�?.11	�?.65
150	�?.87	�?.27

下载: 导出CSV

�?nbsp; 3各国标准中原木材积计算对毓�/p>

Table 3.Comparison of log volume calculation in standards of different countries

国家 Country	计算公式 Calculation formula	实际材积 Actual volume/m3	计算材积 Measured volume/m3	百分比误� Percentage error/%
中国 China	$V = 0.8L{(d + 0.5L)^2}/10\;000$	4.89	4.67	�?.50%
美国 America	$V = ({d^2} - 3d)L/5$	5.38	4.80	�?0.78%
日本 Japan	$V = {d^2}L/10\;000$	4.65	4.43	�?.73%
俄罗� Russia	� None	4.37	4.11	�?.95%
注：V为原木材积，中国、日本的单位为m3，美国的单位为板英尺（bf），1 bf = 2.357 94 × 10�?m3：�i>L为原木检尺长，中国、日本的单位为m，美国单位为英尺（ft）；d为原木检尺径，中国、日本的单位为cm，美国的单位为英寸（in）。Notes:Vis the log volume, the unit of China and Japan is m3, that of American is board foot (bf), 1bf = 2.357 94 × 10�?m3;Lis the log gauge length, the Chinese and Japanese units are m, and the American unit is feet (ft);dis the log gauge diameter, the Chinese and Japanese units are cm, and the American unit is inch (in).

下载: 导出CSV

[2]Arnold R J, Luo J Z, Clarke B. Trials of cold tolerant eucalypt species in cooler regions of south central China[R]. Canberra: Australian Centre for International Agricultural Research, 2004. [3]中华人民共和国国家统计局. 中国统计年鉴(2020)[M]. 北京: 中国统计出版�? 2020.

National Bureau of Statistics of the People’s Republic of China. China statistical yearbook (2020)[M]. Beijing: China Statistics Press, 2020. [4]华蓓, 曹圃, 黄汝�? 基于计算机视觉的原木材积检测方法研究[J]. 河南科技学院学报(自然科学�?, 2022, 50(2): 64�?9.

Hua B, Cao P, Huang R W. Research of log volume measuring method based on computer vision technology[J]. Journal of Henan Institute of Science and Technology (Natural Science Edition), 2022, 50(2): 64�?9. [5]陈广�? 张强, 陈梅�? �? 双目视觉的原木径级快速检测算法[J]. 北京交通大学学�? 2018, 42(2): 22�?0. doi:10.11860/j.issn.1673-0291.2018.02.004

Chen G H, Zhang Q, Chen M Q, et al. Rapid detection algorithms for log diameter classes based on binocular vision[J]. Journal of Beijing Jiaotong University, 2018, 42(2): 22�?0. doi:10.11860/j.issn.1673-0291.2018.02.004 [6]Keck C, Schoedel R. Reference measurement of roundwood by fringe projection[J]. Forest Products Journal, 2021, 71(4): 352�?61. doi:10.13073/FPJ-D-21-00024 [7]唐浩, 王克�? 李晓�? �? 基于色差聚类的原木图像端面检测与统计[J]. 计量学报, 2020, 41(6): 682�?88. doi:10.3969/j.issn.1000-1158.2020.06.09

Tang H, Wang K J, Li X Y, et al. Logs end detection and statistics by color difference clustering[J]. Acta Metrologica Sinica, 2020, 41(6): 682�?88. doi:10.3969/j.issn.1000-1158.2020.06.09 [8]Tang H, Wang K, Gu J C, et al. Application of SSD framework model in detection of logs end[J]. Journal of Physics: Conference Series, 2020, 1486(7): 072051. doi:10.1088/1742-6596/1486/7/072051 [9]蔡瑞�? 林培�? 林耀�? �? 基于改进 YOLOv4-Tiny 的成捆原木端面检测算法[J]. 电视技�? 2021, 45(9): 92�?9.

Cai R X, Lin P J, Lin Y H, et al. A detection approach for bundled log ends based on an improved YOLOv4-Tiny network[J]. Video Engineering, 2021, 45(9): 92�?9. [10]林耀�? 赵洪�? 杨泽�? �? 结合深度学习与Hough变换的等长原木材积检测系统[J]. 林业工程学报, 2021, 6(1): 136�?42.

Lin Y H, Zhao H L, Yang Z C, et al. An equal length log volume inspection system using deep-learning and Hough transformation[J]. Journal of Forestry Engineering, 2021, 6(1): 136�?42. [11]Antti R, Heikki K, Markku T, et al. Electrical impedance and image analysis methods in detecting and measuring Scots pine heartwood from a log end during tree harvesting[J]. Computers and Electronics in Agriculture, 2020, 177(1): 105690. [12]Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]//O’Conner L. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 779�?88. [13]Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]//O’Conner L. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 6517�?525. [14]Redmon J, Farhadi A. YOLOV3: an incremental improvement[C]//O’Conner L. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 1�?. [15]Liu W, Angueloy D, Erhan D, et al. SSD: single shot multibox detector[C]//Leibe B, Matas J, Sebe N, et al. Proceedings of the European Conference on Computer Vision. Berlin: Springer, 2016: 21?37. [16]Girshick R, Donahue J, Darrelland T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//O’Conner L. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Columbus: IEEE, 2014: 580�?87. [17]Girshick R. Fast R-CNN[C]//O’Conner L. Proceedings of the IEEE International Conference on Computer Vision. Santiago: IEEE, 2015: 1440�?448. [18]Ren S, He K, Girshick R, et al. Faster R-CNN: towards realtime object detection with region proposal networks[C]//Kyoung M L. IEEE Transactions on Pattern Analysis and Machine Intelligence. New York: IEEE, 2017: 1137�?149. [19]He K M, Gkioxari G, Dollár P, et al. Mask R-CNN[C]//O’Conner L. Proceedings of the IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 2980�?988. [20]Andrew F, Maurizio P, Robert B F. Direct least square fitting of ellipses[C]//Kyoung M L. IEEE transactions on pattern analysis and machine intelligence. New York: IEEE, 1999: 476�?80. [21]Huang Z J, Huang L C, Gong Y C, et al. Mask scoring R-CNN[C]//O’Conner L. IEEE CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach:IEEE, 2019: 6402�?411. [22]何鸿�? 林晓�? 王林, �? 中国与俄罗斯、美国、日本原木检验标准比较研究[J]. 世界林业研究, 2021, 34(2): 96�?00. doi:10.13348/j.cnki.sjlyyj.2020.0100.y

He H Y, Lin X L, Wang L, et al. Comparative study of log inspection standards among China, Russia, the United States of America and Japan[J]. World Forestry Research, 2021, 34(2): 96�?00. doi:10.13348/j.cnki.sjlyyj.2020.0100.y