文献类型：期刊文献

中文题名：基于YOLOv4的结构用锯材表面缺陷识别

英文题名：Research on surface defect recognition of structural sawn timber using YOLOv4

作者：王勇[1] 张伟[1,2] 高锐[1,3] 金征[1]

第一作者：王勇

机构：[1]国家林业和草原局北京林业机械研究所,北京100029;[2]中国林业科学研究院林业新技术研究所,北京100091;[3]福建省林业科学研究院,福州350012

年份：2021

卷号：6

期号：4

起止页码：120-126

中文期刊名：林业工程学报

外文期刊名：Journal of Forestry Engineering

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金(31670721);中央级公益性科研院所基本科研业务费专项资金(CAFYBB2019MB006);国家林业和草原局林业科学技术推广项目([2019]35)。

语种：中文

中文关键词：深度学习;YOLOv4;锯材表面质量;表面缺陷;质量评价

外文关键词：deep learning;YOLOv4;surface quality of sawn timber;surface defects;quality evaluation

分类号：TP391

摘要：结构用锯材在使用之前进行表面质量评价、分级,对于提高木材的综合利用率具有重要作用。综合利用机器视觉技术和深度学习方法,选取国内常用的云杉结构用锯材作为研究对象,通过工业相机采集结构用锯材表面主要缺陷(节子、虫眼、裂纹),并对锯材主要缺陷进行数字化评价分析。先通过自主搭建的机器视觉图像采集装置,采集100块结构锯材正反面表面图像,共获取表面缺陷图像1 450张,其中活节缺陷图像550张、死节缺陷图像320张、裂纹缺陷图像295张、虫眼缺陷图像285张;随后搭建基于YOLOv4的深度学习缺陷检测识别框架,对缺陷图像中80%的图像进行训练,剩余20%用于测试。试验结果表明,基于YOLOv4的深度学习缺陷检测识别框架,能有效识别并准确定位锯材表面缺陷的类型和位置,平均识别率96.7%,其中活节缺陷识别率100%、死节缺陷识别率97.5%、裂纹缺陷识别率90%、虫眼缺陷识别率96.7%,可满足生产应用需求。
With the increasing shortage of timber resources in the world, how to use timber resources efficiently has already affected the country’s sustainable development plan. The surface quality evaluation and classification of structural sawn timber before use play an important role in improving the comprehensive utilization of wood. This study integrates machine vision technology and deep learning methods and selects domestic commonly used spruce structure sawn timber as a research example. The main defects(knots, wormholes, and cracks) on the surface of sawn timber for structural use are collected by industrial cameras, and the main defects of the sawn timber are digitally evaluated and analyzed. First of all, through the self-built machine vision image acquisition device, the front and back surface images of 100 pieces of structural sawn timber were collected, and a total of 1 450 surface defect images were obtained, including 550 sound knot defect images, 320 unsound knot defect images, 295 crack defect images, and 285 insect defect images. Then a deep learning defect detection and recognition framework is built using YOLOv4, in which, 80% of the defect images are used for training and the remaining 20% for testing. Through experimental tests, it is found that the YOLOv4 algorithm has good convergence properties for small sample data sets and can quickly reduce the loss function value to a small range, and the training accuracy can also be stabilized in a high accuracy range. The test results show that the deep learning defect detection and recognition framework using YOLOv4 can effectively identify and accurately locate the type and location of sawn timber surface defects, with an average recognition rate of 96.7%. Among them, the recognition rate of sound knot defects is 100%, the recognition rate of unsound knot defects is 97.5%, the recognition rate of crack defects is 90%, the recognition rate of insect defects is 96.7%, and the average recognition rate of four types of defects is 96.7%, which meets the requirements of industrial applications.