Research On Visual Detection Technology For Screen Printed Defect Of Lithium Batteries

The quality of lithium battery screen printed is one of the important production indices,and the detection efficiency based on manual visual detection failed to meet the needs of mass production by printing enterprises.The traditional machine vision detection technology could not adapt to the slight deviation phenomenon in the transmission printing process of lithium battery,which results in the decrease of the accuracy of screen printed defect detection.At present,the application of the advantages of deep learning in machine vision detection technology is an important research direction for the intelligentization of printed industry system.In this paper,the lithium battery screen printed defects are taken as the research object,and the image preprocessing,image registration and defect segmentation involved in the visual detection of printed defects are taken as the entry point.Besides,key algorithm theory and technology research are introduced.The main research contents are as follows:(1)Due to the deficiency of over smooth edge information of the traditional image filtering algorithm,an improved weighted guided filtering algorithm based on Do G operator is proposed.Firstly,the Do G operator is introduced to detect the edge of the guided image.And the edge perception weight is obtained.Finally,the adaptive adjustment of penalty term in guided filtering is realized.The image filtering is evaluated based on standard deviation,contrast and signal to noise ratio.According to the experimental results,the proposed algorithm outperformed the traditional image filtering algorithms.In this part,the Do G operator edge detection and adaptive adjustment of penalty term are mainly studied.It can maintain edge information and achieve better denoising effect.(2)Image registration.Differences exist in the spatial positions between the collected lithium battery screen printed images and the template images.The point features detection by AKAZE algorithm showed good robustness but low real-time performance.Therefore,an improved AKAZE image registration algorithm for lithium battery screen printed is proposed.Firstly,the AKAZE feature detection algorithm and BEBLID feature descriptor are combined to extract feature.The method of Brute Force matching is used to extract candidate matching pairs.Then,the GMS algorithm is used to remove the wrong matching pairs between the real image and the template image.Finally,the parameters of the transformation model are estimated by RANSAC algorithm.Bilinear interpolation is employed to complete the final registration.In this method,the performance of different feature extractions and feature matching algorithms in the lithium battery screen printed image registration are researched.Compared with the traditional registration algorithm,the average registration time of the proposed algorithm can reach 0.757 seconds.Besides,the average matching accuracy is improved by 6.42%,with the localization error controlled within 0.8 pixels.(3)Defect segmentation.The defects in the screen printed images of lithium batteries are complex and various,and some of which show low contrast.A two-channel improved U-Net segmentation network is proposed for defect detection.Firstly,the standard file is used as the template image,and the input of the network together with the real image.Secondly,Dense Net121 is introduced to combine high and low level feature information effectively in UNet network.Finally,Focal Loss function is adopted to solve the problem of unbalanced pixel category distribution.In this method,the segmentation performance of deep learning in the lithium battery screen printed defect detection are studied.It is compared with three network models in terms of change detection.The F1-score can reach 0.952,and the average segmentation accuracy of the test phase is improved by 1%.The average detection time was0.047 seconds.The printed defects less than 0.2mm can be accurately segmented.Finally,the detection algorithm was tested and its performance was evaluated.The results show that the performance of detection accuracy and real-time meet the requirements of engineering application.