基于YOLOv5-LW模型的木材表面缺陷检测方法

Detection of wood surface defects using YOLOv5-LW model

为解决人工检测缺陷的不足以及传统视觉检测算法对木材表面细小缺陷漏检率高、检测速度慢、难于部署嵌入式设备等问题,并提高企业木材利用率,以深度学习模型为基础,提出了一种轻量级木板缺陷检测方法YOLOv5-LW。通过重构S 3 Net(ShuffleNetV2+Stem+SPPF)网络模型作为骨干网络,极大减少了模型的参数量和计算量,弥补轻量化后的损失精确度;骨干网络结合ECA注意力机制,可提升网络对关键信息的聚焦能力,并引入特征融合网络MBiFPN,减少特征损失,丰富局部和细节特征,提高细小缺陷的检测能力。本研究将准确率、检测速度、参数量、浮点运算量4个数值作为模型性能的评价指标,通过自制木材缺陷数据集训练,得到8组试验数据,并对模型进行增强前后的对比分析。试验结果表明:改进后的YOLOv5-LW模型相比改进前的原始模型,精确度达到92.8%,在参数量上减少了27.78%,计算量上压缩了40.51%,检测推理速度提升了10.16%,模型对死节、活节、节夹缝、裂缝识别精确率分别为91.8%,87.8%,96.8%和94.9%,对死节和裂痕2种细小缺陷检测精度也有所提升,分别为0.2%和1.6%,且识别效果优于其他6种经典检测模型。因此,本研究所提出的模型更适合部署到嵌入式设备进行木材缺陷实时检测。

In order to solve the problems of manual defect detection such as low efficiency,high cost,and high detection error rate of small defects on the surface of wood,and difficulty in deploying embedded devices by the traditional visual detection algorithms,a deep machine learning model,i.e.,deep learning network model YOLOv5-LW,was used in this study to achieve the rapid and accurate detection of different defects in the wood processing process.By reconstructing the S 3 Net(ShuffleNetV2+Stem+SPPF)network model as the backbone network,the number of parameters and computation time of the model were greatly reduced and the loss of accuracy after lightweighting was deduced.The combination of the backbone network and the ECA attention mechanism improved the network's capability to focus on key information.The feature fusion network MBiFPN was introduced to reduce the feature loss,enrich the local and detail features,and improve the detection capability of fine defects.In this study,the four values of accuracy,detection speed,number of parameters,and floating-point computation were used as the evaluation indexes of the model's performance,and eight sets of experimental data were obtained through the training of the homemade wood defects dataset,and the comparative analysis before and after the enhancement of the model was performed.The experimental results demonstrated that the improved YOLOv5-LW model achieved an accuracy of 92.8%compared with the original model before the improvement,reduced the number of parameters by 27.78%,compressed the computation time by 40.51%,and improved the speed of detection inference by 10.16%.The model accuracy rates for the identification of dead knots,live knots,knots sandwich and cracks were 91.8%,87.8%,96.8%and 94.9%,respectively.The detection accuracies of two types of small defects,dead knots and cracks,were also improved,showing 0.2%and 1.6%,respectively.The recognition effect was better than the other six types of classical detection models,which improved the recognition capability of small wood defects and reduced the errors of detection.Therefore,the model proposed in this study was more suitable to be deployed to embedded devices for real-time detection of wood defects.