Research On Product Packaging Defect Detection System Based On Machine Vision

The production process of dairy products mainly includes milk collection,sterilization,fermentation,filling and packaging.Due to factors such as equipment failure and abnormal raw materials,it is inevitable that there will be defects on the surface of product packaging to varying degrees.The detection of surface defects on product outer packaging is the focus of quality control in the packaging section during the production process.Therefore,it is of great practical significance to carry out research on product packaging surface defect detection.At present,many factories still use manual methods for the detection of packaging defects,which leads to the problems of high labor costs and unstable detection accuracy,which in turn brings the risk of quality complaints to the dairy product market.With the rapid development of machine vision technology,in order to solve the defects of manual detection,this thesis applies machine vision to the production process of dairy products,and studies a dairy packaging defect detection algorithm based on deep learning.Around this theme,the main research work of this thesis is as follows:(1)In view of the development status of defect detection and the environment and needs of the research object in this thesis,dairy packaging detection models based on Faster R-CNN,YOLOv3 and SSD were constructed respectively,and detection experiments based on appearance,straw and sealing defects were carried out.After comparative analysis of the results,the YOLOv3 algorithm with the highest comprehensive performance was selected for improvement and optimization.(2)Aiming at the low detection accuracy of the YOLOv3 network for the slight deformation of the packaging appearance and the inaccurate positioning of some prediction boxes,the following improvement methods are given by analyzing the Darknet53 structure and the calculation principle of the YOLOv3 loss function:Introduce GIo U Loss to replace the original network MSE Loss,Increase the 104*104prediction layer and increase the number of a priori boxes.The experimental results show that the m AP of the improved YOLOv3 algorithm model is increased by 2.83%,the average defect AP is increased by 5.8%,and the comprehensive performance is optimized.(3)In view of the improved YOLOv3 defect detection algorithm and actual requirements in this thesis,an operating system software based on Py Qt5 is designed,including login,registration,picture detection,video detection and real-time detection,which realizes real-time monitoring of dairy product packaging defects,expands its practical application.Through the research on the detection of dairy packaging defects,the defect identification effect has been improved,the m AP has reached 93.7%,and the average defect AP has reached 88.06%.Preliminary exploration of real-time automatic detection of dairy packaging defects in production lines.