Research On Defect Classification Method Of Printed Image Based On Machine Learning

There are often different kinds of printed matter defects in the process of printing and their harms are also different.Some defects are generated with randomness,just need to be marked out;And the generation of some defects is cyclical,belong to serious defects,once there must be shut down for corresponding elimination,otherwise it will cause serious material waste and product quality problems.Therefore,the classification of printing defects has important practical application value.Due to the complex and changeable forms and colors of printing defects produced in production practice,the forms of printing defects of the same type are greatly different.In addition to dealing with the differences among different types,classification methods should also consider the challenge of excessive changes within the same type of printing defects,which makes the research on this topic of academic research value.At present,there are many researches on the detection of printing defects at home and abroad,but few on the classification of printing defects.This subject through the cooperation with domestic printing enterprises,according to the actual requirements of enterprises,using machine learning method,from many aspects of the printing defect classification method.The main research contents are as follows:(1)establishment of printing defect data set.By cooperating with domestic printing manufacturers,the project obtained various sample images of printing defects in actual production.According to the actual production requirements,the data of printing defects were divided into four categories:dark defect,bright defect,cutter wire and color slant.Different modal images of various types of printing defects,including standard template images,original defect images,binarization defect images,difference defect images and defect images with background removed,were established,which laid a foundation for the research on the classification method of printing defects.(2)the application of traditional machine learning method in the classification of printing defects is studied.Firstly,the difference defect image is binarized,and the center position and minimum bounding rectangle of the defect image are determined.Secondly,the morphological features of the defect image after binarization were extracted,including defect area,defect roundness,defect axial ratio and average circumference of defect connected domain,etc.,to reflect its morphological information.Then,according to the location information determined on the defect image after binarization,the brightness features,including the mean brightness of the defect and the variance of the brightness change in the defect area,are extracted from the original defect image to reflect the brightness information.Finally,all features were normalized,and defect classification experiments were carried out on halcon platform with SVM and MLP respectively.The average classification accuracy reached 90%and 89%,respectively.(3)research on the classification method of printing defects based on deep learning was carried out.First study different modal data on classification accuracy,the influence of the deep learning method using data including image and the original defects after binarization of the defect images,the results show that with the background of the original defects had a greater influence on the figure of classification accuracy,thus put forward a kind of weed out the background image printing defect image data processing method,improves the classification accuracy of the results.On this basis,a printing defect classification method based on double-channel CNN is designed and compared with the experimental results of VGG16 and ResNet18.The results show that the printing defect classification method based on double-channel CNN designed in this paper has a high classification accuracy.(4)A deep neural network operating framework based on TensorFlow was built on Ubuntu 16.04,and a printing defect classification ex perimental system was designed by combining with the UI development environment Tkinter of python.This system can be used to classify and test printed defect data with different network models,and the classification results can be visualized.