Research On Surface Defect Detection Of Polarizer Based On Deep Learning

As an indispensable part of liquid crystal displays,polarizers have a wide range of application prospects in the display field.Their quality will directly affect the display effect of liquid crystal displays.Surface defect detection of polarizers has become an important part of the manufacturing process.At present,relying on manual detection methods can no longer meet actual production needs.Traditional detection methods have poor generalization ability to extract features,while deep learning can learn independently from a large amount of data,and the extracted features are more representative.This paper uses the target detection algorithm based on deep learning to complete the surface defect detection of the polarizer.The main research content includes the following aspects:(1)Research the relevant theories of deep learning detection algorithms,and introduce the classic convolutional neural network models in detail.These models are used to extract the surface defects of the polarizer.The selection of Res Net-101 through experimental comparison can better learn its surface existence independently.Defects;The target detection network represented by R-CNN and YOLO is analyzed in detail,and the efficiency and accuracy of defect detection are comprehensively considered.Faster R-CNN is selected as the basic framework to establish a detection model,and the target is given Detect relevant evaluation indicators.(2)Research on the detection method of polarizer surface defects based on Faster RCNN.Aiming at the problem of inability to identify small defect sizes during the feature extraction process in the Faster R-CNN algorithm,it is introduced in the feature extraction network Res Net-101 The Feature Pyramid Network(FPN)constitutes the residual-feature pyramid network;in view of the ROI pooling operation in the network,there will be two quantization rounding resulting in missing pixels,a multi-level ROI pooling structure is designed to replace the traditional rough ROI pooling layer,And through experiments to determine the division scheme of the multi-level ROI pooling structure;the anchor frame scheme in the original RPN network is not suitable for the established polarizer defect data set,and the clustering algorithm k-means++ is used for optimization and improvement;The improved Faster R-CNN network is used for polarizer surface defect detection.Through experiments and other algorithms for comparative analysis,the improved Faster R-CNN network can guarantee accuracy while also having a good detection speed.(3)Research on the surface defect detection method of polarizer based on Mask R-CNN.Mask R-CNN is a network based on Faster R-CNN.The main difference is that Mask RCNN adds parallel branches to each ROI.Carrying out instance segmentation,mainly focusing on adjusting the number of neurons in the "Head" part,adjusting the feature extraction network and removing the mask branch three strategies to optimize and improve,and combine experiments to analyze their impact on network performance,and use different algorithms to detect time The accuracy of the prediction results of the bounding box and the eye mask determines that Mask R-CNN-512 can achieve the best performance in detection speed and accuracy.The improved Faster R-CNN and Mask R-CNN and the original YOLOv3 are polarized the test images are compared in experiments,and the performance of these three methods in the surface defect detection task of polarizers is analyzed according to the number of detections,missed detection rate and false detection rate.Finally,the performance of the improved Mask R-CNN is shown through experimental data.Better than the other two methods.