Research And Implementation Of Deep Learning Based Defect Detection Algorithm For Industrial CT Workpiece Perspective View

In 2013,after Germany proposed Industry 4.0 at the Hannover Messe,China also proposed 《Made in China 2025》 in 2015,and the concept of intelligent manufacturing began to take root in people’s hearts;at the same time,the continuous development of China’s economy and society and the improvement of people’s quality of life,the requirements for production quality are also improving year by year,but subject to the limitations of the working environment,production conditions and other factors,the workpiece produced may However,due to the working environment,production conditions and other factors,the produced workpiece may have defects,especially the workpiece with high precision requirements,but also need to use industrial CT to detect its internal,and the defect detection based on industrial CT image is still mainly manual detection,this way is more subjective,relying on manual experience,more prone to error.With the development of computer vision,deep learning provides a new solution for industrial production,AI empowerment of traditional CT detection,avoiding the impact of detection results because of subjectivity and improving the accuracy of detection,which is of great significance to intelligent manufacturing.In this thesis,we implement industrial CT image defect detection by deep learning methods based on the internal defect scenario of industrial CT images,and propose a multiscale perceptual field semantic segmentation network with the following main work and research.1.Using labelme software,label the collected industrial CT images,and clean the data of the labeled images,correct the wrongly labeled images,complete the data set,and lay the data foundation for the experiment.2.For the influence of artifacts and noise in the CT imaging process,which leads to the problem of blurred defect boundaries and low contrast in the scanned CT images,the method of homomorphic filtering is used to pre-process the images and enhance the contrast between the defects and the background to facilitate detection.3.For the problem of misalignment defects in industrial CT images of ceramic chips,this thesis uses a deep learning classification network approach to detect them using VGG16,Res Net50 and Inception V3,respectively.Through experimental comparison,Inception V3 is selected,which achieves 97.7% accuracy and 15.7 FPS in misalignment recognition.4.For the bubble and crack defects of ceramic chips and the crack defects of ceramic workpieces,this thesis proposes a RPPM-UNet semantic segmentation network model based on UNet,with the following two main points: an RPPM multi-scale perceptual field module is designed to maximize the extraction of network information and strengthen the inter-pixel connection;the Focal loss function is invoked to strengthen the positive sample defect weighting factor.The improved network achieves 74.86% and 76.62% Miou and 10.31 and11.28 FPS on ceramic chip and ceramic workpiece CT datasets,respectively,and the detection accuracy of CT defects is improved compared with the original UNet.5.In order to speed up the inference of the model,in this thesis,after quantizing the model using NVIDIA Jetson Xavier NX deployment platform,the FPS of classification network reaches 29.08,and the FPS of ceramic chip and ceramic artifact in semantic segmentation network reaches 17.89 and 20.94 FPS.6.Using Py Qt as a platform,we design an industrial CT defect detection system,embed the network model into the detection system,implement the defect detection of images in the system,and visualize the results.In summary,this thesis integrates image processing,classification networks,and semantic segmentation to achieve industrial CT defect detection based on a deep learning approach.Through the study of industrial CT defect detection,from model design to the construction of the detection system,it has certain theoretical value and practical significance.