Research On Circuit Defect Detection Of TFT-LCD Based On Deep Learning

Thin film transistor liquid crystal display(TFT-LCD),as the mainstream device of modern display equipment,has complex manufacturing technology and composition structure,especially in its interior,which contains a large number of compact and small size electronic circuits.Traditional human eye detection or machine vision detection has been unable to meet the requirements of high precision detection.The detection technology based on deep learning can fully grasp the law of the image and extract the essential features of the target defects by training the massive sample data,so as to achieve the purpose of recognition and classification,and greatly improve the detection precision.Based on the actual collection of TFT-LCD circuit defect images,this paper constructs an effective algorithm model combining the advantages of machine vision and deep learning technology,and studies the effect of deep neural network in the field of TFT-LCD defect detection.The main research contents and results are as follows:(1)Through the study of the principle of image acquisition in the automatic optical detection(AOI)system,the influence of different optical imaging devices,such as light source,lens and CCD camera,on the image quality is analyzed and compared.For the common defects in TFT-LCD,such as panel scratches,conductive particle bonding and the problem of line end location,target segmentation and recognition are performed using a combination of machine vision algorithms such as threshold segmentation,morphological processing,and template matching.(2)TFT-LCD electronic circuit is easy to arise fine scratches,breakages,dirty,foreign objects and other defects,which present blurred edge and low contrast.It is difficult to detect with the traditional method of human eyes and image processing.In this paper,a method of combining histogram equalization with convolution neural network is proposed.Enhance the gray level of image through preprocessing first,and then send into the constructed deep convolution neural network for feature extraction and classification.Training and testing on the actual collection of TFT-LCD circuit defect image data sets,the effects of different dropout parameters and batch sizes on the accuracy are analyzed in detail.The results show that the network model has the best classification effect under the parameters of dropout ratio of 0.5 and the batch size of 128,and achieve an accuracy of 94.6%,which 10.85%higher than the CNN of non image preprocessing and 18.08%higher than traditional SURF feature matching algorithm.(3)Accurate recognition and location of defects are the requirements for high automatic detection of TFT-LCD.On the basis of completing the task of defect classification,a region based neural network is proposed to identify and locate the target defects.Constructing a multi-layer regional proposed network to improve the structure of original Faster R-CNN network,and enhance the selection ability of the image candidate area,thereby deal with the problem that the TFT-LCD circuit defects is too small to generate precise candidate area.At the same time,a variety of basic network structures are designed to study the influence of different neural network depth and convolution kernel size on the detection results,and compare them with different methods.To further expand the original sample data,build a data set containing 43200 images of six types of TFT-LCD circuit defects,and make reasonable division and annotation processing.The results of training and testing on the data set show that the network structure with 16 layers depth and 3x3 size convolution kernels has the best detection effect,and the detection system can recognize and locate each kind of TFT-LCD border circuit defects in one image simultaneously within 0.12s and achieve an accuracy of 94.6%.