【作者】 高杰；

【导师】 童水光； 童哲铭；

【作者基本信息】 浙江大学 ， 机械设计及理论， 2020， 硕士

【摘要】 近年来,人工智能产业以及图像识别技术飞速发展,在工业生产中的应用日益广泛,为现代化智能制造关键问题不断提供新的解决方案。轮毂制造属于庞大的汽车零部件制造业,具有产量不一、形状复杂、规格型号繁多的特点,对使用图像识别技术解决轮毂识别问题的实际应用提出了越来越高的要求。因此本文面向工业生产中含有复杂产线背景的轮毂图像识别问题,以图像识别技术为研究对象,设计了一套多工况下的轮毂类型识别系统。首先,本文分别从传统分类技术和现代分类技术两个角度阐述了轮毂图像识别技术的课题背景与研究现状,提出了本课题的研究目标,并据此对本文主要研究内容与章节安排进行阐述。其次,结合硬件实验环境详细介绍了基于图像识别技术的图像预处理方法的基本原理。针对轮毂图像,研究并优化了基本的图像去噪及对比度增强方法;基于轮毂的图像特征提取技术,研究了图像的阈值分割以及形态学算法及实现。然后在图像预处理算法的基础上,研究并实现基于特征工程的轮毂图像提取技术。针对毛刺量不大但含有强背景干扰的轮毂图像识别问题,本文基于Canny边缘检测算法,结合直方图分割技术改进图像差分算法以及采用图像分段线性变换技术,实现了复杂背景下的轮毂特征提取,解决了包括轮毂外圆,轮辐数量以及内孔数量在内的特征提取问题。针对毛刺过大且传统特征提取方法不适用的轮毂图像识别问题,基于尺度不变特征(SIFT)的局部特征提取与匹配算法构建最近邻分类器,并引入在线学习策略,对随机采集的6种轮型120张图像,实现了单张识别1.5秒,识别率100%的优秀结果,解决了基于图像处理等传统方法无法识别大毛刺的轮毂问题。接着结合现代深度学习方法,通过图像预处理算法完成了轮毂数据集的构建与增强,基于改进的卷积神经网络ResNet-12模型,借鉴集成学习策略,采用迁移优化算法进行模型训练,实现了在19种类型68622张轮毂图像上99.95%以上的高识别率,且单张识别时间仅消耗20毫秒不到,解决了轮毂的高速高精度识别问题。最后为了适应生产实际需要,基于Windows平台开发轮毂图像(视频)实时识别系统,实现了轮毂视觉识别技术的集成。更多还原

【Abstract】 In recent years,the artificial intelligence industry and image recognition technology have developed rapidly,and their applications in industrial production have become increasingly widespread,providing new solutions to the key issues of modern intelligent manufacturing.The manufacturing of wheels belongs to the huge automobile parts manufacturing industry,which has the characteristics of different output,complicated shapes,and various specifications.This puts forward higher and higher requirements for the practical application of image recognition technology to solve the problem of wheel recognition.Therefore,this paper is aimed at the image recognition of wheel with complex production line background in industrial production.Basedon the image recognition technology,a set of wheel type recognition system under multiple operating conditions is designed.Firstly,this article expounds the background and current research status of wheel image recognition technology from two perspectives: traditional classification technology and modern classification technology.The research objectives of this topic are put forward,as well asthe main research content and chapter arrangement of this article.Secondly,the basic principle of image preprocessing method based on image recognition technology is introduced in detail in combination with hardware experimental environment.For the wheel image,the basic image denoising and contrast enhancement methods were studied and optimized.And based on the wheel’s image feature extraction technology,the threshold segmentation,morphological algorithm and implementation of the image were studied.Then this paperstudiesand implement the wheel image extraction technology based on feature engineeringandimage preprocessing algorithm.Aiming at the problem of wheel image recognition with small glitches and strong background interference,this paper improves the image difference algorithm based on histogram segmentation technology and adopts the image segmentation linear transformation technology to achieve wheel feature extraction under complex background,and finally solves the problem of feature extraction including the outer circle of the wheel,the number of spokes and the number of inner holes.Aiming at the disadvantage of poor real-time performance of traditional recognition methods,this article combined with modern deep learning methods to complete the construction and enhancement of the wheel data set through image preprocessing.Its main idea is to use the improved convolutional neural network ResNet-12 model,borrow ensemble learning strategies and use migration optimization algorithms for model training.This model achieves a high recognition rate of more than 99.95% on 19 types of 68,622 wheel images,and the recognition time of a single sheet takes less than 20 milliseconds,which solves the problem of high-speed and highprecision recognition of wheels.Finally,in order to meet the actual needs of production,a wheel image(video)real-time recognition system was developed based on the Windows platform,and the integration of wheel visual recognition technology was realized.更多还原