Research On Surface Defect Detection Technology Of Printing Roller Based On Machine Vision

As the basic industry of national economy,printing industry plays an important role in social service,cultural development and information dissemination.Printing machinery in the printing industry has a pivotal position,printing roller as the core components of printing machinery,in the production and processing process may appear a variety of defects,its quality directly affect the quality of printing products.At present,the domestic roller defect detection classification is generally mainly manual,the main problem of manual detection is low detection efficiency,high production costs,quality control is not strict.Aiming at these practical problems,improving the production efficiency,using the method of machine vision and machine learning to realize the detection of the surface defects of the printing roller has become an inevitable trend.Taking the surface defect image of the drum of the printing press as the research object,this paper makes a deep study on the preprocessing,defect detection and defect classification algorithm of the surface defect image of the roller,and completes the construction of the system architecture.The main research is as follows:(1)This paper analyzes the demand and method of surface defect detection of roller in domestic printing press,and introduces the method based on visual and deep learning into the surface defect detection of roller cylinder for the first time,and promotes the development of intelligent detection method in this industry.(2)In the stage of image preprocessing,a uniform light processing algorithm is proposed to eliminate the influence of bright and dark stripes on image quality in view of the problem of shading and dark stripes collected from images.Aiming at the influence of dust,noise and other interference on the image quality of roller in industrial production site,several filtering algorithms are compared with experimental analysis,and the Gaussian filtering algorithm is finally selected,and the experimental results Gaussian filter can remove the noise of the image of the roller surface well.At the same time,several classical image segmentation methods are analyzed.(3)The detection algorithm of roller surface defect is analyzed and studied.By analyzing the advantages and disadvantages of two defect detection algorithms based on image difference and template matching,a roller surface defect detection algorithm based on adaptive Canny operator is proposed,and the threshold value of edge extraction is determined by OSTU method according to pixel amplitude characteristics,and the threshold value is used to replace the low threshold in Canny operator.Then the defect region is connected by morphological open operation,and finally the defect location is found accurately according to the eight neighborhood boundary tracking algorithm,and the required characteristics are obtained.(4)The classification algorithm of convolution Neural network is analyzed and studied.In order to meet the requirements of the system "detection-Classification" single job time,this paper adopts the classical LeNet-5 model as the classifier,which has the advantages of simple structure and easier convergence in training.At the same time,the sliding model,L2 and sparse DropConnect methods are introduced,which have a good performance in preventing over-fitting and improving the generalization ability of the network.