基于循环卷积神经网络的缺陷图像评判分级系统  

Defect Image Evaluation and Classification System Based on Cyclic Convolution Neural Network

作　　者：李长春[1] 赵卫东[1] LI Chang-chun;ZHAO Wei-dong(Chuzhou Vocational Technology College,Chuzhou 239000,China)

机构地区：[1]滁州职业技术学院,安徽滁州239000

出　　处：《长春师范大学学报》2021年第4期38-42,共5页Journal of Changchun Normal University

基　　金：安徽省教育厅重点研究项目“基于人工智能技术的‘平安校园’智能安防系统研究与设计”(KJ2019A1136);安徽省教育厅重点研究项目“基于‘互联网’的多维度智能高校实验室管理系统的研究”(KJ2020A0992);安徽省教育厅重点研究项目“大数据环境下基于数据挖掘的教学质量监控与评价系统”(KJ2020A0998);安徽省质量工程精品开放课程项目“数据库应用技术”(2017kfk144);安徽省质量工程校企合作示范实训中心项目“软件开发实训中心”(2019xqsxzx31);滁州职业技术学院校级质量工程精品资源共享课程项目“网络设备配置与管理”(2018jpkc004)。

摘　　要：为了解决当前实际场景缺陷图像难识别、缺陷图像库数据庞大而导致标记工作难度大的问题,本文基于自主开发软件标注算法,对大数据缺陷样本库进行半自动标记,并在自主开发软件内完成标记。本文方法基于循环卷积神经网络框架,对标记结果进行学习训练,形成精准识别机制,使其不依赖第三方标注软件。首先,采集大数据缺陷图像,为缺陷识别做好数据训练准备。然后,结合传统视觉检测技术中的图像对比和图像阈值分割,实现初期检测。最后,比较多种深度学习框架的特性,开发了深度学习模型,并将其集成到本文系统中,建立深度神经网络缺陷识别算法。实验测试结果显示,本文系统具有更高的缺陷识别精度与鲁棒性,可为智能缺陷识别设备奠定算法基础。In order to solve the problem that it is difficult to identify the defect image in the actual scene and the huge data in the defect image database,this study based on the self-developed software annotation scheme,semi-automatic marking of the big data defect sample library is carried out,and the marking is completed in the self-developed software,and the labeling is not dependent on the third-party annotation software.Based on the framework of the cyclic convolution neural network,the labeling is completed results to carry out learning and training to form a precise recognition mechanism.Firstly,the big data defect image is collected to prepare data training for defect recognition.Then,combined with image contrast and image threshold segmentation in traditional visual inspection technology,the initial detection is realized,which can meet the needs of equipment detection.Finally,the advantages and disadvantages of various deep learning frameworks are compared.Based on these frameworks,the deep learning model is developed and integrated into the system to establish the defect recognition mechanism of deep neural network.The experimental results show that the system is conducive to the landing of high-precision defect recognition system,and lays the algorithm foundation for intelligent defect recognition equipment.

关 键 词：半自动标记 缺陷图像 图像对比 循环卷积神经网络 深度学习框架 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]