【作者】 李建明；

【导师】 杨挺； 张志斌；

【作者基本信息】 天津大学 ， 电气工程， 2019， 硕士

【摘要】 缺陷检测是工业生产中产品质量控制的关键环节,也是目前工业生产实现全自动化的一个瓶颈。基于人工的产品缺陷检测方法存在检测效率低、成本高、容易误检漏检等缺点,已经无法适应现代高速自动化的工业生产要求。而基于传统机器视觉的缺陷检测方法,虽然可以在一定程度上弥补人工检测的缺点,但是其检测设备往往结构复杂、价格昂贵,只能应用于对特定目标的检测。当检测目标改变时,需要由具备专业知识的人重新进行复杂的特征提取和特征选择过程,由于人工设计的特征不能有效覆盖缺陷的全部特征,所以在检测效率和准确率上受到了限制。同时在复杂的工业环境和复杂的检测背景下,传统的机器视觉检测方法往往也受到了限制。本文研究了缺陷检测的发展现状,针对目前工业缺陷检测中面临的问题,在样本数据较少的情况下,提出了一种基于深度学习的Inception-V3图像分类算法和YOLO-V3目标检测算法相结合的工业自动化缺陷检测方法,避免了传统机器视觉缺陷检测算法复杂的特征提取过程:首先将待检测图片通过YOLO-V3目标检测模型获取“感兴趣”区域,然后对这些区域依次进行数据增强后通过Inception-V3缺陷识别模型,对获取的识别结果取平均值后和设定的阈值比较,得到缺陷识别结果。本文通过采集现场图片,制作实验数据集,采用数据增强、迁移学习等技术分别训练了YOLO-V3缺陷区域检测模型和Inception-V3缺陷识别模型,然后通过实验验证了本文提出的缺陷检测方法的有效性,相比只应用图像分类算法的检测方法具有更高的准确率和稳定性,同时该方法能应用在复杂的检测环境和检测背景下,在多目标检测中体现优势。本文针对提出的缺陷检测方法设计了一整套模块化的缺陷检测系统,包括数据采集模块、目标检测模块、图像识别模块、报警模块、数据存储模块。模块间通过MQTT、Redis的发布订阅机制和Modbus TCP协议实现通讯或数据共享。缺陷识别模型通过Tensor Flow Serving进行部署,支持g RPC调用和热部署,方便模型管理。该系统硬件成本低,部署和扩展操作简单,可应用于大规模的工业自动化检测场景中,具有很好的应用价值,对中小企业实现自动化缺陷检测,提高生产效率具有重要的意义。更多还原

【Abstract】 Defect detection is not only the key link of product quality control in industrial production,but also the bottleneck of realizing full automation in industrial production.At present,the artificial product defect detection method has some disadvantages,such as low detection efficiency,high cost and easy false detection,which can not meet the production requirements of high-speed automation in industry.And the defect detection method that based on traditional machine vision,to some extent,although can make up for the shortcoming of artificial detection,its detection equipment is often complicated in structure and expensive in price,and can only be applied to the detection of certain goals.When the detection target changed,a person with professional knowledge is needed to complex the feature extraction and feature selection process again.Due to the fact that the characteristics of the artificial design can not effectively cover the all features that make one defect,so this defect detection method is limited in efficiency and accuracy.At the same time in the complex industrial environment and complex detection background of the application scene,the traditional machine vision detection methods are also limited.This paper studies the development status of defect detection,In view of the problems faced in current industrial defect detection field,an industrial automated defect detection method based on the combination of Inception-V3 image classification algorithm and YOLO-V3 target detection algorithm is proposed under the condition of few sample data,which avoids the complex feature extraction process of traditional machine vision defect detection algorithm: First,the image to be tested is put into the YOLO-V3 target detection model to obtain the "areas of interest",and then these areas are successively enhanced with data augmentation,and then the Inception-V3 defect recognition model is adopted to obtain the final defect recognition result by taking the average value of the obtained recognition results and comparing with the set threshold value.In this paper,we trained the YOLO-V3 target detection model and Inception-V3 defect recognition model respectively through collecting pictures,making the experiment data sets,using data augmentation and transfer learning technology.Then,the validity of the defect detection method proposed in this paper is verified through experiments.Compared with the detection method that only applies image classification algorithm,the detection method has higher accuracy and stability.Meanwhile,the method can be applied in complex detection environment and detection background,showing advantages in multi-object detection.In this paper,a complete set of modular defect detection system is designed for the proposed defect detection method,including data acquisition module,image detection module,image recognition module,alarm module and data storage module.Communication and data sharing between modules is realized through Modbus TCP protocol and the publish and subscribe mechanism of MQTT and Redis.The recognition model is deployed via Tensor Flow Serving which supports g RPC invocation and hot deployment to facilitate model management.Not only the hardware cost is low,but also the operation of deployment and expansion is simple of this system and can be applied to large-scale industrial automation application scenarios.It is of great significance to small and medium-sized enterprises to realize automated defect detection and improve their production efficiency.更多还原