| 基于深度学习的人造板表面缺陷检测研究 |
| |
| 引用本文： | 魏智锋,肖书浩,蒋国璋,伍世虔,程国飞.基于深度学习的人造板表面缺陷检测研究[J].林产工业,2021(2):21-26. |
| |
| 作者姓名： | 魏智锋 肖书浩 蒋国璋 伍世虔 程国飞 |
| |
| 作者单位： | 武汉科技大学机械自动化学院冶金装备及其控制教育部重点实验室;武昌首义学院机械与自动化学院;中山火炬职业技术学院 |
| |
| 基金项目： | 基于机器视觉的人造板非平面缺陷检测(51874217);广东省普通高校青年创新人才类项目(2018GkQNCX049);中山市社会公益科技研究项目(2018B1113);国家自然科学基金(31670561)。 |
| |
| 摘 要： | 缺陷识别是人造板检测的重要环节,目前大多采用人工检测方法。将一种轻量级的深层神经网络MobileNet与SSD算法结合,使用Inception网络附加到多个特征映射上,构建SSD-MobileNet算法模型用于人造板的缺陷检测,以提高区分不同缺陷的能力。从人造板工厂生产现场获取主要包括粗刨花、水印、砂痕、杂物、胶斑5种缺陷类型的表面缺陷图,制成一个包含3216张人造板表面缺陷图像的数据集。利用该数据集对SSD-MoblieNet模型进行训练、测试,并与其他特征提取网络(ResNet18、VoVNet39、ESPNetV2)的检测精度和检测速度的影响结果进行对比,发现其检测速度最快达到75帧/s,相对其他特征提取网络的平均精度均值提升2.26%~3.52%。该研究为实现人造板表面实时在线检测提供良好的技术支撑。
|
| 关 键 词： | 人造板 表面缺陷 SSD-MobileNet 卷积神经网络 深度学习 检测 |
Research on Surface Defect Detection of Wood-based Panels Based on Deep Learning |
| |
| WEI Zhi-feng,XIAO Shu-hao,JIANG Guo-zhang,WU Shi-qian,CHENG Guo-fei.Research on Surface Defect Detection of Wood-based Panels Based on Deep Learning[J].China Forest Products Industry,2021(2):21-26. |
| |
| Authors: | WEI Zhi-feng XIAO Shu-hao JIANG Guo-zhang WU Shi-qian CHENG Guo-fei |
| |
| Institution: | (Key Laboratory of Metallurgical Equipment and Control of Ministry of Education,School of Mechanical Automation,Wuhan University of Science and Technology,Wuhan 430081,Hubei,P.R.China;School of Machinery and Automation,Wuchang Shouyi University,Wuhan 430068,Hubei,P.R.China;Zhongshan Torch Vocational and Technical College,Zhongshan 528436,Guangdong,P.R.China) |
| |
| Abstract: | The defect identification of wood-based panels is an important part of its detection.At present,manual detection is mostly used.To improve the ability of distinguishing different defects,a lightweight deep neural network,MobileNet,was combined with the SSD algorithm,and Inception was used to attach to multiple feature maps to construct an SSD-MobileNet algorithm model for defect detection of wood-based panels.Various surface defect maps which mainly included five types of defects:rough shavings,watermarks,sand marks,debris,and rubber blocks,were obtained from the production site of the wood-based panels factory,and were made into a 3216-based image data set of surface defects of the wood-based panels.By using this data set to train and test the SSD-MoblieNet model,and comparing the results of other feature extraction networks(ResNet18,VoVNet39,ESPNetV2)on the detection accuracy and detection speed,it was found that the speed reached 75 frames per second at the fastest.Compared with the mAP of the feature extraction network,it increased by 2.26%~3.52%.This study would provide good technical support for real-time online detection of the surface of the wood-based panels. |
| |
| Keywords: | Wood-based panels Surface defects SSD-MobileNet Convolutional neural network Deep learning Defect detection |
| 本文献已被 CNKI 维普 等数据库收录！ |
|
