基于改进YOLOv5的纸病检测方法  

Paper Defect Detection Method Based on Improved YOLOv5

作　　者：张开生[1] 关凯凯 ZHANG Kaisheng;GUAN Kaikai(School of Electrical and Control Engineering,Shaanxi University of Science&Technology,Xi’an,Shaanxi Province,710021)

机构地区：[1]陕西科技大学电气与控制工程学院,陕西西安710021

出　　处：《中国造纸》2022年第10期79-86,共8页China Pulp & Paper

基　　金：陕西省榆林市2020年科技计划项目(CXY-2020-090)。

摘　　要：针对传统纸病检测算法中纸病特征提取困难、实时性差的问题,提出一种改进YOLOv5的纸病检测方法。该方法首先在批量归一化模块的首尾部分添加居中和缩放校准,形成更稳定的纸病有效特征分布;其次在骨干网络中添加坐标注意力机制,增强骨干网络的纸病特征提取能力;最后选用CIoU_loss作为边界框回归的损失函数,实现高精度的定位。实验结果表明,改进后的算法平均精度达99.02%,实时检测速度达41.58帧/s,相较于现有的基于CNN纸病分类算法,检测精度与检测速度都有较大的提升,且改进后的算法对光源的依赖程度更低,能对各类纸病实现精准辨识。Aiming at the problems of difficulty in extracting paper defect features and poor real time in current paper defect detection algorithms,an improved YOLOv5 paper defect detection method was proposed in this study.The method was firstly adding centering and scaling calibration to the head and tail of the batch normalization module to form a more stable effective feature distribution of paper defects.Secondly,a coordinate attention was added to the backbone network to enhance the paper defect feature extraction capability of the backbone network.Finally,CIoU_loss was selected as the loss function of bounding box regression to achieve high-precision positioning.The experimental results showed that the average accuracy of the improved algorithm reached 99.02%,the real-time detection speed reached 41.58 frames/s.Compared with the existing CNN-based paper defect classification algorithm,the detection accuracy and detection speed were greatly improved.The improved algorithm was less dependent on the light source and could accurately identify various paper defects.

关 键 词：纸病检测 YOLOv5 批量归一化模块 坐标注意力机制 损失函数 

分 类 号：TS736[轻工技术与工程—制浆造纸工程] TP391[自动化与计算机技术—计算机应用技术]