【作者】 李勇；

【导师】 王杰； 胡明春；

【作者基本信息】 四川大学 ， 机械工程（专业学位）， 2022， 硕士

【摘要】 化学纤维在日常生活中运用十分广泛,其品质的好坏将直接影响到化纤织物的外观与性能。化学纤维在生产出来以后为便于运输通常会缠绕在纸筒上,形成丝筒或丝饼。在生产过程中由于多种因素的影响会导致次品的产生,在装箱之前进行丝筒表面质量检测是十分重要的。目前运用的检测方法通常为人工通过肉眼观察判断是否存在缺陷,这种检测方法不仅效率低下而且由于人的主观因素影响导致评判标准波动,容易导致检测准确度不佳。化纤丝筒产量特别巨大,设计一套针对性的丝筒表面缺陷检测系统显得极为迫切,运用自动化检测达到降本增效的目的。本研究分析了目前已有的丝筒表面缺陷检测方法存在的不足,以及丝筒表面缺陷的特征,设计了一套有针对性的图像采集系统。不仅采集了常见的表面异物、绊丝等明显缺陷,还采集了目前鲜有研究的包含色泽缺陷的图像,对色泽缺陷进行检测。制作完成缺陷物体检测数据集和色泽分类数据集后运用计算机视觉图像处理技术,经过实验分析最终采用了YOLO v5将丝筒表面图像的缺陷区域框选出来。在对接头区域进行接头计数任务时,提出了一种适用性更好的接头检测算法。色泽图像分类任务里对Bi FPN结构进行了改进,提出了一种新的通道注意力机制模型,在Bi FPN的单个节点里施加通道维度的注意力进行特征筛选,以提升模型的判别能力。设计了一个色泽图像分类模型,该模型可以分为三个部分,分别是特征提取模块、特征融合模块以及分类器模块。在特征融合模块里运用本研究中提出的改进后的Bi FPN结构进行深层特征与浅层特征的融合。从多层次多角度进行对比实验,通过大量实验证明了模型的有效性,模型对于丝筒表面色泽分类具有强大的实力。最终形成了Efficient Net作为特征提取模型,改进的Bi FPN模型作为特征融合部分并结合分类器的丝筒表面色泽分类模型。该模型不仅准确率最高,达到了99.06%,而且参数量也比较少,单张图像推理速度约为0.11s/张,具备很强的运用潜力。最后设计了一个交互界面,将硬件图像采集部分和软件算法部分相结合,输出最终分级结果,形成了一套丝筒表面缺陷检测及分级系统。更多还原

【Abstract】 Chemical fibers are widely used in daily life,and their quality will directly affect the appearance and performance of chemical fiber fabrics.After the chemical fibers are produced,they are usually wound on paper tubes to form silk cylinders or silk packages for easy transportation.Due to the influence of various factors during the production process,it is important to inspect the surface quality of the cones before packing.The current inspection method is usually manual observation to determine the presence of defects,which is not only inefficient but also subjective to human factors that lead to fluctuations in the judging criteria,which can easily lead to poor inspection accuracy.The production volume of chemical fiber cylinder is particularly huge,so it is urgent to design a set of targeted surface defect detection system for the chemical fiber yarn cylinder to achieve cost reduction and efficiency improvement by using automated inspection.In this study,we analyzed the shortcomings of the existing surface defect detection methods and the characteristics of the surface defects of the fiber cylinder,and designed a targeted image acquisition system.Not only common obvious defects such as surface foreign objects and tripped filaments are captured,but also images containing color and luster defects,which have rarely been studied so far,are captured to detect color and luster defects.After creating the defective object detection dataset and luster classification dataset using computer vision image processing techniques,YOLO v5 was finally used to frame out the defective areas of the surface images of the silk cylinder after experimental analysis.For the joint counting task in the joint area,a better joint detection algorithm was proposed for applicability.The BiFPN structure is improved in the color and luster image classification task,and a new channel attention mechanism model is proposed to impose channel dimensional attention in a single node of Bi FPN for feature screening to improve the discriminative power of the model.A color and luster image classification model is designed,which can be divided into three parts,namely the feature extraction module,the feature fusion module,and the classifier module.The improved Bi FPN structure proposed in this study is applied in the feature fusion module to fuse deep features with shallow features.Comparative experiments are conducted from multiple levels and perspectives,and the effectiveness of the model is proved through a large number of experiments,and the model has strong strength for silk cylinder surface color and luster classification.Finally,Efficient Net is formed as the feature extraction model,and the improved Bi FPN model is used as the feature fusion part and combined with the classifier for silk cylinder surface color and luster classification model.The model not only has the highest accuracy of 99.06%,but also has a relatively small number of parameters,with an inference speed of about 0.11s/image for a single image,which has a strong potential for application.Finally,an interactive interface is designed,which combines the hardware image acquisition part with the software algorithm part to output the final classification result,forming a set of wire cylinder surface defect detection and classification system.更多还原