Research On Defect Detection System Of Industrial Products Based On Machine Vision

Defect detection is a critical step in the industrial manufacturing process to control the quality of products,and it is significant to realize automated defect detection to promote the development of "smart" manufacturing industry.In recent years,deep learning algorithms have been widely used in the field of defect detection and have achieved good detection results,but most of the existing research is only for a specific detection scenario,and high performance requirements for the equipment running the detection program,which does not have good transferability and promotion value.In this dissertation,by summarizing the commonalities of defect detection scenarios,we design and implement a migratable and highly adaptable defect detection application framework using a neural network-based machine vision approach,and propose two lightweight improvement methods to make it run smoothly in low-resource embedded platforms and achieve high accuracy and low latency detection results.The main work of this dissertation as follows:(1)A deep learning defect detection framework based on machine vision is designed and implemented to address the common problems in defect detection tasks.The framework uses the YOLOv5s model to identify specific target locations and classes in the image of the detected product to make defect judgments,and mitigates the difficulty of defect data set acquisition by means of data augmentation and transfer learning techniques,and finally transposes its detection part to an embedded platform for running tests.In addition,the overall framework is modularly packaged in this dissertation so that it has good migratability.(2)In order to improve the operation of the defect detection framework in the embedded platform,two lightweight improvement methods are proposed for the YOLOv5s object detection network used in the framework.One is to analyze the composition of the feature extraction network,and select some of the convolutional layers to be improved into lightweight convolutional layers with enhanced receptive fields to achieve the goal of reducing the network complexity while ensuring accuracy.The second is to redesign the detection head part of the network,remove the anchor mechanism used in the original network,and modify the prediction frame representation method and the positive sample matching method in the training process to enhance the model performance while avoiding the step of fitting the anchor frame to the data set during network training.After experimental tests,the network models with the application of both improved methods can improve the model inference speed in embedded devices.(3)Self-picked images of cable components and annotated to create a dataset of cable assembly components,combined with the proposed and improved inspection framework for the detailed design of the automated defect detection system for cable assembly components,and finally deployed the system to the embedded platform Jetson Nano for testing.The test results show that the machine vision-based defect detection solution for industrial products proposed in this dissertation can meet the actual operational requirements and has certain engineering application value.