【摘要】 针对木条表面死结和活结缺陷在检测过程中定位困难、平均识别精确度较低、检测速度较慢的问题,在分析木结缺陷特点和改进深度学习YOLOv3模型的基础上,研究其应用于改善木结缺陷检测时的精确度和速度。首先,对活结缺陷图像进行数据扩增,以解决类别不平衡问题。然后,改进k-means++算法,提升木结缺陷目标框的维度聚类效果,得到更合适的初始目标框个数与尺寸;通过缩减YOLOv3中多尺度检测网络、改进损失函数,以减少检测时间和提高目标识别精确度。最后,对木结缺陷进行拼接得出位置坐标。试验结果表明,较改进前YOLOv3算法,mAP值提升7.47%,检测速度提高35%;较Faster R-CNN算法mAP值提升11.68%,检测速度提高约15倍,改进后模型能精确地检测出死结和活结缺陷。因此,在后续研究中,可考虑以YOLOv3算法作为检测木结缺陷模型,进一步改进YOLOv3网络,以提高检测实时性和精确度。更多还原

【Abstract】 To address the problem of difficult positioning,low mean average precision and slow detection speed in the process of detecting dead knots and live knots on the wood surface,based on the analysis of wood knot defects characteristics and improved deep learning network YOLOv3 model,a study that it is applied to improve the accuracy and speed of detecting wood knot defects is proposed.Firstly,the data amplification of live knot defect image is carried out to solve the category imbalance problem.Then,k-means++ algorithm is improved to enhance the dimensional clustering effect of the target boxes of wood knot defects,and the more appropriate number and size of initial target boxes are obtained.The multi-scale detection network of YOLOv3 is reduced and the loss function is modified to reduce detection time and improve target recognition accuracy.Finally,the wood knot defects are spliced to obtain the position coordinates.The experimental results show that compared with the original YOLOv3 algorithm,the mAP value is increased by 7.47%,and the detection speed is increased by 35%.Compared with the Faster R-CNN algorithm,the mAP value is increased by 11.68%,and the detection speed is increased by 15 times approximately.The improved YOLOv3 can accurately detect dead knot and live knot defects.Therefore,in the future research,the YOLOv3 algorithm can be considered as a model for detecting wood knot defects,and the YOLOv3 network can be further improved to improve the real-time and accuracy of detection.更多还原