摘要
锯材表面缺陷检测是锯材加工的重要环节,随着深度学习和卷积神经网络的发展,越来越多的目标检测算法被引入并应用于锯材表面缺陷的检测。然而,目前的锯材表面缺陷检测算法仍存在检测速度慢、检测精度低等问题。为了解决这些问题,设计了一种基于YOLOv5的锯材表面缺陷检测算法,针对该算法设计了可重参数化的主干网络,改进了聚类算法和损失函数,使用了多种数据增强方式并且设计了解耦的检测头。将改进后的算法在NVIDIA RTX3090 GPU上进行训练和测试,在本研究使用的橡胶木和松木锯材表面缺陷数据集上的平均精度均值(mAP)分别达到了98.02%和98.37%,推理时间分别为6.79和6.58 ms。将本研究改进的算法与改进前的算法以及Faster R-CNN、YOLOv3等常用算法进行对比,结果表明,本研究改进的YOLOv5在速度和精度方面都有显著提高。此外,还将本研究使用的橡胶木和松木数据集进行了融合,本研究改进的算法在融合后数据集上的mAP达到了94.88%,推理时间为6.89 ms,相比对比算法仍有明显优势。最后,对相关研究成果进行了对比分析,结果表明本研究改进的算法在检测速度上也具有较大优势,验证了该算法的有效性。
Surface defect detection in sawn lumber is crucial for maintaining high-quality standards in the lumber industry.By detecting surface defects,the industry can ensure that only high-quality lumber reaches the market.This contributes to better overall product quality and customer satisfaction.With the development of science and technology,the demand for higher quality sawn timber has increased.Although the development of deep learning and convolution neural networks has allowed more objective algorithms to be used for the detection of surface defects,the current detection algorithms are slow and poor accuracy.Therefore,improving the accuracy and detection speed of sawn timber surface defect detection algorithms is a crucial challenge.In this study,an improved object detection model was designed based on YOLOv5 with a re-parameterized backbone network to improve the detection speed.A clustering algorithm for YOLOv5 was also proposed to obtain anchors that was suitable for the two datasets.In addition,this study improved the loss function by introducing the SIoU loss function to enhance the accuracy of the prediction boxes and used several data augmentation methods to improve the detection accuracy of the model.Finally,the detection header was decoupled to calculate the prediction box coordinates,confidence,and category probabilities,respectively.Using the NVIDIA RTX3090 graphics processing unit,the improved algorithm achieved a mean accuracy percentage(mAP)of 98.02%and 98.37%on the two sawn lumber surface defect datasets,which were 1.73%and 1.72%higher compared to the original algorithm.The detection times were 6.79 ms and 6.58 ms,which were reduced compared to the speeds of 1.74 s and 1.44 s using the original algorithm.The improved algorithm in this paper was compared with the previous algorithm and common algorithms such as Faster R-CNN and YOLOv3,and the results showed that the improved YOLOv5 in this study had significantly improved speed and accuracy.In addition,the two datasets used in this study were fused,and the improved algorithm achieved a mAP of 94.88%on the fused dataset,with a detection time of 6.89 ms,which had a significant advantage over other comparative algorithms.Finally,the related research results were discussed,revealing that the improved algorithm in this study had a great advantage in detection speed,which also verified the effectiveness of the improved algorithm.
作者
杨昊
张茹
王钰圣
赵园园
毕立岩
任世学
王伟
YANG Hao;ZHANG Ru;WANG Yusheng;ZHAO Yuanyuan;BI Liyan;REN Shixue;WANG Wei(Engineering Research Center of Advanced Wooden Materials,Ministry of Education,Harbin 150040,China;College of Material Science and Engineering,Northeast Forestry University,Harbin 150040,China;Chenxu Automation Equipment Co.Ltd.,Guangzhou 510000,China;College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China)
出处
《林业工程学报》
CSCD
北大核心
2024年第5期134-143,共10页
Journal of Forestry Engineering
基金
黑龙江省自然科学基金(LH2019C009)。
