基于Halcon深度学习的高分子薄膜表面缺陷检测  

Surface Defect Detection of Polymer Film Based on Halcon Deep Learning

作　　者：邓海云 陈新辉 Deng Haiyun;Chen Xinhui(College of Technology,Shantou University,Shantou,Guangdong 515063,China)

机构地区：[1]汕头大学工学院,广东汕头515063

出　　处：《机电工程技术》2024年第7期153-157,共5页Mechanical & Electrical Engineering Technology

基　　金：广东省科技专项资金项目(STKJ2023056)。

摘　　要：详细阐述了高分子薄膜表面的各类缺陷类型和特征及其可能对高分子薄膜的性能和功能上造成的破坏,旨在说明在生产过程中对高分子薄膜表面缺陷进行检测的必要性。针对高分子薄膜表面缺陷检测中应用深度学习大模型训练而存在各缺陷类别样本数量较少、缺陷特征变化不足而导致检测精度低等问题,应用了一种基于Halcon全局上下文异常值检测算法,详细阐述了该检测算法的GC-AD Combined训练模型,其是一种融合了Faster-RCNN和Auto-Encoder的网络结构。在较为有限的高分子薄膜合格的图像和含有缺陷的图像数据基础上,实验设计了采用GC-AD Combined网络模型对高分子薄膜图像样本进行训练,待训练完成后对多组薄膜图像数据进行推理测试,同时设计了与CNN、Mobile-net以及VGG等网络模型检测的对比实验。实验结果显示,所运用的GC-AD Combined网络模型的平均准确率高达98.07%,能够满足吹塑薄膜表面缺陷检测任务,检测精度大幅度优于其他3种网络模型,进一步说明了应用基于Halcon全局上下文异常值检测算法的GC-AD Combined模型完成高分子薄膜表面质量检测任务的有效性和优越性。The types and characteristics of various defects on the surface of polymer films,as well as the potential damage they may cause to the performance and functionality of polymer films are elaborated.The aim is to demonstrate the necessity of detecting surface defects in polymer films during the production process.In response to the problems of low detection accuracy caused by the small number of samples in various defect categories and insufficient changes in defect features in the application of deep learning large model training in surface defect detection of polymer thin films,a global context outlier detection algorithm based on Halcon is applied,and the GC-AD Combined training model of the detection algorithm is elaborated.It is a network structure that integrates Faster RCNN and Auto Encoder.On the basis of relatively limited qualified images and image data containing defects of polymer films,an experimental design is carried out using the GC-AD Combined network model to train polymer film image samples.After the training completed,multiple sets of film image data are subjected to inference testing.At the same time,comparative experiments are designed with network models such as CNN,Mobile-net and VGG for detection.The experimental results show that the average accuracy of the GC-AD Combined network model used is as high as 98.07%,which can meet the surface defect detection task of blow molded films.The detection accuracy is significantly better than the other three network models,further demonstrating the effectiveness and superiority of using the GC-AD Combined model based on the Halcon global upper and lower anomaly detection algorithm to complete the surface quality detection task of polymer films.

关 键 词：深度学习 HALCON 图像处理 缺陷检测 机器视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]