| With the rapid development of science and technology,the requirements for advanced production technology in the industrial field are gradually increasing.Our country proposed the development strategy of "Made in China 2025",focusing on technological innovation and artificial intelligence technology to enhance the core competitiveness of the national manufacturing industry.The intelligent inspection technology based on machine vision has important and extensive applications in positioning inspection and component surface quality inspection in industrial automation manufacturing.Machine vision positioning detection technology can accurately calculate the pose coordinate information of the object to be measured,so as to realize automatic positioning guidance and automatic assembly of the object.In the surface quality inspection of objects,the traditional manual visual inspection method has the problems of low efficiency,strong subjectivity,and inability to work continuously for a long time.The machine vision intelligent detection technology combined with optical imaging and image processing technology has the advantages of accuracy,high efficiency,and long continuous working time with no contact.It is the most ideal method for component surface detection.With the explosive development of artificial intelligence technology,machine vision combined with deep learning can improve the detection accuracy comparing with traditional image processing algorithms.Currently it has been the most promising development direction of machine vision detection technology.The main research areas of this thesis are the real detection applications based on machine vision intelligent detection techniques,focusing on the key technical fields such as camera models,imaging system designs,imaging process algorithms in feature extraction,feature recognition,feature matching,and feature classification.There're many problems unresolved in the real applications of pose estimation of objects and high-precision quantitative detection of surface defect of optical elements,which are the key research issues in this thesis.The main contents of this thesis include:Aiming at the three-dimensional positioning detection scene where only the single plane image of the object can be obtained by monocular camera,a set of overall positioning detection system from imaging acquisition,feature extraction,positioning calculation and mechanical system coordinate guidance is established.An imaging system composed of an array camera and supporting light sources is established,using ORB to extract features and obtain its description operator.The algorithm based on spatial constraints and random sampling and consensus(SCRSAC)is proposed to obtain the inner matching point pairs with high accuracy and strong robustnes,and the pose of objects based on the PnP algorithm model can be obtained.The key point of machine vision positioning technology lies in the establishment of the positioning model and the location extraction and matching algorithm of object features based on the imaging system.For object surface defect detection,the key lies in the segmentation extraction and classification algorithm of defect features based on the high-quality images obtained by imaging systemIn the detection of surface defects,frosted glass is mainly used as the detection object.Unlike ultra-smooth surface glass,the scattering characteristics of the frosted glass surface are more complex,and the dark field-based imaging system cannot meet the detection requirements.Therefore,a dual-channel linear array scanning imaging system(CPBIS)based on coaxial parallel illumination and bright-field imaging was established.Line scan cameras were used to obtain high-resolution glass images.The distortion introduced by the lens and the background non-uniformity caused by light source non-uniformity were analyzed.A non-uniformity correction method based on statistical background estimation was proposed to improve the uniformity of the image background.Aiming at the requirements of glass surface defect detection,defect segmentation algorithm based on reconstruction of local background model was proposed,and a neighbor defect merge algorithm was proposed based on the neighborhood distribution principle.For the classification problem of glass surface defects,the classification data and sample enhancement of the defect data obtained during the detection process were established.Based on the inspiration of Inception and ResNet networks,an improved balanced residual network structure(Modified Inception-Res-Net,MIRNet)is proposed,which reduces the depth of the network and increases the width of the network.The dataset of 17600 images containing 8 types of defects was established,which was used to verify the validity of the proposed model.The weak imaging defects such as weak scratches,light discoloration and light dents on the glass surface are more difficult to detect in the traditional algorithm.The image segmentation algorithm based on symmetric convolutional neural network(Symmetrical Net,SymNet)was proposed.By labeling the defect data collected during the inspection process at the pixel level,a dataset containing more than 30,000 defects and background samples was established.The defects featrues can be obtained in various dimensions,achieving better performance than the traditional glgorithmsThe experiments were set up to verify the validation of the proposed systems.The pose estimation system is applied to the automatic battery replacement of electric vehicles,reaching the positioning error within 1mm.The glass detection system is applied to the detection of frosted mobile phone back glass,reaching the classification accuracy performance of more than 94.2%.The recall rate of segmentation detection for weak defects reaches 95.3%,and the accuracy rate reached 91.8%. |
PDF Full Text Request