Research On Defect Detection Method Of Backlight Mura Based On Deep Learning

As one of the important components in liquid crystal display(LCD),the backlight module is developing towards high resolution,light weight and low power consumption.As the size of the backlight module increases and the thickness decreases,the probability of various defects(MURA)in the backlight module increases greatly.But for the traditional backlight defect detection,most enterprises use manual detection,which has strong subjectivity,low efficiency,lack of quantitative indicators and other shortcomings,resulting in unsatisfactory detection rate,low production efficiency and other problems.Therefore,it is very necessary to design a backlight defect detection system with rapidity,robustness and conformity with human eye evaluation standards.The Mura defect in backlight has the characteristics of fuzzy boundary and low contrast,and few Mura defect samples are produced in the process of backlight production,which brings a challenge to defect detection.Based on the characteristics of Mura defects,this thesis studies the backlight Mura defect detection method based on deep learning.The specific research contents are as follows:(1)Based on the backlight region extraction and data expansion method,the backlight Mura defect data set was constructed.In view of the problem that the backlight image collected in the field contains the background such as the operation table and may produce distortion,the corrected screen area is obtained by using Canny algorithm,Hough line fitting and perspective transformation.To solve the problem of poor generalization of the training model caused by the small number of Mura defect images,two methods of advadance generation network and dynamic data enhancement of simulated field conditions were used to expand the training set data.(2)AResNet classification defect detection method based on sliding window and weighted feature fusion is proposed.The backlight image is divided based on the sliding window idea,and the divided sub-image is input into the network to judge whether there is a defect or not.The defect is located by the position of the sub-image containing the defect.In order to solve the problem that the feature information of Mura defect edge mostly exists in the shallow features,which leads to the low detection rate of the single feature image extracted by ResNet,the residual network was used to extract features at different levels by combining the idea of feature pyramid and weighted fusion,and the weighted fusion of features was carried out by ReLU function and fast normalization.The efficiency and stability of fusion are improved while the feature semantic information is improved.Experimental results show that the proposed algorithm can effectively improve the detection rate of Mura defects.(3)A method for CenterNet defect detection based on attention mechanism is proposed.In order to solve the problem that the ResNet classification defect detection method is not accurate,the end-to-end CenterNet target detection network is used.The weighted feature fusion ResNet method is used to extract features from backlight images in backbone network.Aiming at the problem that low contrast Mura defect is difficult to detect under the condition of uneven brightness,attention mechanism was introduced to improve the attention of the network to the target defect.Experimental results show that the proposed algorithm can accurately locate the defect while guaranteeing the defect detection rate of Mura.