Research And Application Of Key Technologies For Online Semantic Segmentation-based Machine Vision Inspection And Identification

Machine vision inspection technology is one of the key technologies involved in intelligent manufacturing equipment,and is widely used in precision manufacturing production lines,online product quality inspection and other areas of intelligent manufacturing.This doctoral thesis is entitled "Research and Application of Key Technologies for Online Semantic Segmentation-based Machine Vision Inspection and Identification".The thesis focuses on the key technical issues such as online complex background semantic segmentation method,auto deep learning method of semantic segmentation model,and multi-task parallel scheduling method for online machine vision inspection and identification.It is of significance to improve the level of machine vision inspection technology and equipment.From the need for accuracy and real-time on-line machine vision inspection,the thesis discusses domestic and international research progress in deep learning semantic segmentation methods,automatic machine learning methods,and multi-task scheduling methods.The main work of the doctoral thesis includes:(1)Overall architecture design of online semantic segmentation-based machine vision inspection and identification technology.Based on the accuracy and real-time requirements of online machine vision inspection,the overall architecture and detection process of online semantic segmentation-based machine vision detection and identification technology are established.The main indexes include accuracy indexes such as mean intersection over union IoU and mean average precision APIoUT,as well as real-time indexes such as segmentation time Tseg and training time Ttrain.The key technologies of the architecture analysis is points out,mainly including online machine vision semantic segmentation,auto deep learning,and multitask scheduling.Among them,the online complex background semantic segmentation method both improve the accuracy index IoU APIoUT and reduce the segmentation time Tseg.The auto deep learning method helps to reduce the training time Ttrain.The multi-task scheduling method helps to reduce the total segmentation time Tseg under various detection tasks.(2)Research on online complex background semantic segmentation methods for machine vision inspection and identification.The key technologies such as fewer parameter optimization of network model based on atrous convolutional architecture,complex background semantic segmentation based on coder-decoder architecture network,and feature map reuse of full convolutional network were proposed.The conflicts of network semantic segmentation capability-computational overhead,object component semantic response value-semantic mutually exclusive relationship,and network semantic segmentation accuracy-feature reuse degree were resolved systematically.The improved semantic segmentation model has higher accuracy IoU?APIoUT and shorter segmentation time Tseg.And it facilitates in-line machine vision inspection applications.(3)Research on automatic deep learning methods for semantic segmentation machine vision inspection and identification.The transfer learning,hyperparameter optimisation configuration and fast validation techniques for semantic segmentation of machine vision are studied systematically.The conflicts of initial loss reduction of training networks,convergence stability and computational overhead are resolved respectively.The proposed automatic deep learning method for machine vision detection with transfer learning and hyperparameter optimization is characterized by less training generations Ntrain and higher accuracy APIoUT,which is convenient for applications in detection environments where inspection components and features change frequently.(4)Research on multi-task scheduling method for online semantic segmentation machine vision inspection and identification.The technical framework of the multi-task scheduling module for online semantic segmentation machine vision inspection This study establishes.The high-resolution image block scheduling,multi-network multi-image batch scheduling,and inspection and identification collaborative scheduling key technologies are investigated.The conflicts of hardware usage-peak storage overhead,preloaded network storage overheadaverage GPU usage,and scheduling solution time-scheduling optimization time have been solved.The system is able to improve the real-time performance of machine vision inspection with shorter total segmentation time Tseg in the case of high-resolution images,multi-network multi-images and complex inspection.(5)Experimental study of online complex background semantic segmentation machine vision inspection and identification applications.Validate the theory and method of this doctoral thesis in practical application in terms of functionality,accuracy and real-time requirements.The preliminary application of the AK8 bank note anti-counterfeiting identification shows that,it has semi-supervised learning bank note anti-counterfeit feature class,position and shape..The banknote anti-counterfeiting identification accuracy is 100%,identification time is less than is.The application of the MVAQ-2 chassis standard component assembly quality inspection system show that,the system can semi-supervised learning different assembly types,positions and shapes.The MVAQ-2 system has the ability of inspecting different component assembly situation,with a inspection time of no more than 5.3s and an accuracy rate of 100%.The learning time for a single chassis less than 20.2 min.The application of the DAE driving identification system shows that the system can assess whether the test is passed or not according to the operation behaviour during the test.The accuracy of the key functions such as controller gear and manoeuvre behaviour recognition is Pgear=100%,Poperation=98.3%.The total Exam time error ?TDAE,gear shift operation moment error ?Tforklift,and the moment of turn operation error ?Tturn do not exceed ±0.310s,satisfying the needs of the driving examinations.