Research On Defect Detection And Classification Of Light Guide Plate Based On Machine Vision

With rapid economic growth,liquid crystal display(LCD),owing to its advantages in environmental protection,replaces the traditional cathode-ray tube to dominate the display market.The backlight unit(BLU)not only provides the light source to LCD,but also has a direct impact on imaging effect.The thinner entangled BLU has earned the favor of many manufacturers.In recent years,how to improve the excellent characteristics of light guide plate(LGP)of entangled BLU has become the focus of research.To ensure the quality of the mesh hot-stamping process of LGP and determine whether the process meets the requirements,the intactness of mark lines and points is tested.The main work and achievements of the present study are summarized as follows:1.Research on the Traditional Visual Technology1.Research on the Traditional Visual Technology(1)In the traditional visual technology,the machine vision pre-processing algorithm,the feature extraction algorithm and the statistical method are used for exploration and demonstration.(2)By investigating the LGP manufacturers on site and analyzing the characteristics of the tested targets,traditional image algorithms are used to detect some more obvious features like mild contamination and mild bubbles.The mean filter for texture structure of mark lines is improved to implement unidirectional filtering.By relying on a range of computer vision algorithms such as threshold segmentation,morphological method,angle correction,and extreme value extraction of projection,the fitting and difference of high-order polynomials is applied to anticipate the position of mark lines.The standard deviation is finally used to eliminate the interference and determine the position of mark lines.(3)Regarding the detection of LGP mark points,set the threshold based on a priori knowledge for threshold segmentation.The morphological method is used for scaling the position of edge.A mask is created to apply convolution in the ROI region.The mean square error of gray level of pixels around mark points is calculated.The position of mark points is finally determined to visualize the testing results of mark lines and points.2.Deep learning technology(1)In this paper,a mixed convolution neural network is designed,and a Dense Net-BCNN convolution neural network is designed to classify the defects such as air bubbles,no-line and line in the surface of the light guide plate.(2)In this paper,the modified Dens Net network is used as a feature extractor for feature extraction,and the Billinar-CNN fine-grained algorithm is used to improve the attention of the convolution neural network and improve the detection accuracy of the defect.(3)In this paper,the three-layer full connection layer is used to replace the Global Average Pooling algorithm,and the Dense Block structure is modified to avoid over-fitting.We use transform learning method to train Dense Net-BCNN to improve the accuracy of defect detection and reduce the number of parameters and defect detection time.(4)The defect classification of convolution neural network in gray images of low contrast,high texture and high similarity industrial small data sets are discussed.3.proof of algorithm(1)After in practical test,the average accuracy rate is 99.78 percent and the average detection time is 166.5ms,which meets the practical production requirements,in the case of the elimination of serious pollution of the plate,serious edge of the marking line and excessive bubble area.(2)After testing,deep learning technology is obtained with a parameter of1.14 MB,and an average detection time of 40.1ms.The network parameters are 97.2% which is lower than the V2-Res Net-101 network.