Research On Defect Detection Algorithm Of TFT-LCD Based On YOLO

Recently,semiconductor display industry is developing rapidly.Although the automation of panel production is getting higher and higher,most of the anomalies depend on artificial discrimination.The accuracy of artificial discrimination of defects is affected by experience,which is easy to cause capacity loss due to misjudgment.With the rapid development of neural network,the target detection algorithm represented by YOLO is constantly updated.This paper sets up a dataset of TFT-LCD images in the process of array manufacturing,and proposes a set of automatic defect detection scheme based on YOLO(You Only Look Once),which involves three aspects: defect location,defect classification and defect segmentation.The specific research contents include:(1)In terms of defect location and defect classification,in order to solve the problem of poor detection effect of YOLO due to the different size of TFT-LCD defects and similar features among categories,a cascaded algorithm of defect location and classification of TFT-LCD based on YOLO and Res Net is proposed.YOLO is responsible for the location of defects and Res Net is responsible for the classification of defects.Firstly,in order to improve the accuracy of YOLO positioning defects,this paper designs a feature extraction network applied to the backbone of YOLO,and expands multi-scale fusion on the basis of YOLO to strengthen the detection rate of large defects and small defects;Secondly,in order to meet the needs of YOLO for single category training,the loss function is pruned and improved;Finally,in order to strengthen the classification accuracy of similar defect categories,a logical block is used to cascade YOLO and Res Net.In the logic block,the size of the defect image is fixed based on the center point of the positioning frame and then transferred to the classification network to avoid the problem of information loss due to scaling of the defect.In addition,the area weighting method is used to improve the "attention" of large defects.After YOLO is improved,the AP under single category training can reach0.87.(2)In terms of defect segmentation,in order to solve the problem of difficult and expensive data annotation in the training segmentation network,an unsupervised TFTLCD defect segmentation algorithm based on the Generative Adversarial Networks is proposed.Firstly,the attention image is obtained by training the generator of Attention GAN through the dataset with only defect and no defect domain.for the problem of unstable training of the Attention GAN and poor generation of attention images,WGAN-GP is used to optimize the loss function.Secondly,the defect mark image and defect mask image are obtained according to the significance of attention map.Finally,the binary image is obtained by the region growth of binary morphological reconstruction.(3)In order to further solve the problem of poor segmentation due to insufficient training data,a defect segmentation algorithm based on original image comparison is proposed.This method first uses the defect area located by YOLO to make a mask image and then uses it as a convolution kernel,combines convolution and correlation coefficients to find the defect-free area with the largest background correlation with the defect area in the original image,and then subtract the two to get the final segmentation result.Experimental results prove that combining the above two segmentation methods can meet the segmentation requirements of most TFT-LCD defect categories in industrial scenarios.In an experimental environment with an i7-7700 K CPU clocked at 4.2GHZ,a GTX1080 GPU with 8G video memory,and a deep learning framework of Pytorch1.8,the experiment shows that in a single-category experiment,the improved YOLO AP@0.5 is increased by 6%,6%,7.5% respectively compared to YOLOv3,YOLOv3-SPP and YOLOv3-TINY;and in the multi-category experiment,the F1 score of the cascaded defect detection algorithm is increased by 23%,27%,47% compared with YOLOv3,YOLOv3-SPP and YOLOv3-TINY respectively.Based on the two segmentation methods of Generative Adversarial Networks and original image comparison,it can meet the segmentation requirements of most TFT-LCD defect categories in industrial scenarios.