基于改进YOLOv5的玻璃纤维管纱缺陷检测方法  

Detection method of glass fiber tube yarn defect based on improved YOLOv5

作　　者：董振宇 景军锋[1,2] DONG Zhenyu;JING Junfeng(Xi'an Polytechnic University,Xi'an,710600,China;Xi'an Polytechnic University Branch of Shannxi Artificial Intelligence Joint Laboratory,Xi'an,710600,China)

机构地区：[1]西安工程大学,陕西西安710600 [2]陕西省人工智能联合实验室西安工程大学分部,陕西西安710600

出　　处：《棉纺织技术》2023年第12期12-19,共8页Cotton Textile Technology

基　　金：国家自然科学基金项目(62176204);陕西省创新能力支撑计划项目(2021TD-29);陕西省秦创原“科学家+工程师”队伍建设项目(2023KXJ-061);陕西高校青年创新团队项目。

摘　　要：针对玻璃纤维管纱缺陷检测中存在的抗干扰能力差、检测精度低和检测速度慢的问题,提出了一种基于改进YOLOv5的玻璃纤维管纱缺陷检测方法(BY-YOLO)。首先建立了高效重参数网络(ER-Net)作为主干网络对管纱缺陷特征进行优化提取,利用结构重参数化技术和精确金字塔池化模块(R-SPP)提升检测速度,减弱特征噪声信息对检测效果的影响;其次提出了深度注意力路径聚合网络(DA-PANet)作为颈部网络对管纱的多尺度特征进行融合,通过特征增强模块Depth-Mixer和注意力机制模块增强管纱缺陷特征的语义信息,提高模型对多尺度缺陷的检测能力。试验结果表明:该方法能够将玻璃纤维管纱缺陷检测的mAP值提高至94.43%,同时将其检测速度提升到103帧/s。与其他主流的检测模型相比,该研究提出的方法拥有更高的鲁棒性、准确性和实时性。Aiming at the problems of poor anti-interference ability,low detection accuracy and slow detection speed in glass fiber tube yarn defect detection,a glass fiber tube yarn defect detection method based on improved YOLOv5(BY-YOLO)was proposed.Firstly,Efficient Reparameterization Network(ER-Net)was established as the backbone network to optimize the extraction of defective features of tube yarn.And structural reparameterization techniques and Refined Spatial Pyramid Pooling(R-SPP)were used to enhance the detection speed and diminish the influence of noise information of features on the detection effect.Secondly,Depth Attention Path Aggregation Network(DA-PANet)was proposed as a neck network to fuse the multi-scale features of tube yarn.The semantic information of defective features of tube yarn was enhanced by the feature enhancement module Depth-Mixer and the attention mechanism module.The detection capability of the model for multi-scale defects was improved.The experimental results showed that the method was able to improve mAP value of the detection on tube yarn defects to 94.43%.At the same time,the detection speed was increased to 103 flame/s.Compared with other mainstream detection models,the method proposed in this paper had higher robustness,accuracy and real-time performance.

关 键 词：管纱缺陷检测 机器视觉 深度学习 YOLOv5 结构重参数化技术 注意力机制模块 平均精度均值 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]