Defect Detection And Classification Of LCD Panel Based On Deep Learning

With the ultra-high-speed iterative update of electronic products and equipment in recent years,smart devices have increasingly demanded screen LCD panels,and defect detection of LCD panels is an important link affecting screen quality.The current panel defect detection mainly depends on two methods: one is to judge by human eyes with a magnifying glass,which is very dependent on the experience of workers and the waste of labor costs is serious;the other is to use the traditional method of combining computer vision and machine learning.The detection efficiency of both methods is relatively low,can not meet the factory's real-time requirements for defect detection,and the current types of defects become more diverse with the complexity of the panel processing technology,the traditional methods can not meet the accuracy requirements of defect detection.Therefore,this paper takes the defect data in a panel production factory as the research object,and designs two kinds of defect detection algorithms for liquid crystal panels based on deep learning technology,in order to improve the detection efficiency and accuracy of liquid crystal panel defects.This article first analyzes the related theoretical techniques of AMOLED panel defect detection,combines the collected image data,uses flipping,cropping,random defect generation and other methods to expand the data set,preprocesses the data set based on this,and uses the LabelImg tool to collect the image Calibrate the defect type and defect location,and convert the tag data into a usable form.The model migration method is used to fix some parameters of the deep convolutional network,and small sample data is extracted to perform feasibility analysis of Faster RCNN algorithm and RetinaNet algorithm in this experiment to verify its convergence.Secondly,a comparative experiment of Faster RCNN algorithm under different processing conditions is designed.The ROI Align algorithm is used to rewrite the ROI Pooling algorithm of the extracted part of Faster RCNN candidate area.The FPN network structure is compared with the RetinaNet model.The Faster RCNN algorithm runs less efficiently than RetinaNet,but the detection accuracy is about 10% higher than RetinaNet,but both of them are much higher than the current industry average for defect detection.Finally,considering the application of this experiment in actual production scenarios,the model is researched on pruning,weight sharing,and quantification of lightweight.On the basis of maintaining the accuracy of the original network training,the network parameter occupation memory is reduced by nearly 8 times.The experimental results in this paper show that using deep learning to detect defects on LCD panels can not only reduce the empirical error of manual inspection,but also avoid the complex feature construction in traditional visual methods.By comparing the one-stage and two-stage methods,Faster RCNN algorithm is optimized,and the lightweight method is used to compress the model to ensure that the model can be applied on the mobile end and improve the actual production efficiency.