Research On On-line Detection Of TFT-LCD Chips Based On Machine Vision

TFT-LCD(Thin Film Transistor Liquid Crystal Display)chips are commonly applied in various display devices.In recent years,the market demand for its quality is increasing with the rapid development of TFT-LCD.Automatic,efficient and accurate detection of TFT-LCD chips has become a critical segment for manufacturers to guarantee product quality and control product costs.Machine vision based on automatic optical detection has the advantages of high precision and fast speed,which has become the mainstream method of automatic detection at present.Based on machine vision,this paper focuses on the research of accurate,high efficiency and automatic quality detection of TFT-LCD chip manufacture.The main research contents are as follows:(1)Image processing algorithm for conductive particle detection based on gradient method is studied,and a conductive particle detection algorithm based on regional characteristic gradient is proposed.In order to realize the detection of conductive particles,based on the characteristics of differential interference imaging and conductive particles,the pseudo-illuminant direction is estimated by statistical method,and then the candidate particles are obtained.The detection algorithm of conductive particles is designed based on the regional characteristic gradient and the feature of the region near the candidate particle is extracted.Commonly used clustering algorithms are compared and the optimized K-means algorithm is selected to cluster the candidate particles.The recognition rate is 98.61%.In order to meet the real-time performance of industrial detection,the parallel computing based on CUDA is used to accelerate the speed of the algorithm and the time consumption is kept within 2 seconds.(2)The conductive particles detection algorithm based on machine learning is studied,and a data augmentation method for conductive particles is proposed.The imbalance problem of the existing conductive particle dataset is analyzed.The freeform deformation is applied to expand the low-intensity particle dataset and the continuity of the boundary of the deformation area is analyzed.The experimental result show that additional features will not be introduced into the dataset using intensity deformation.The recognition effects of the prediction model based on Mask R-CNN trained by the dataset augmented by simple method and the dataset augmented by intensity deformation proposed in this paper are compared.The experimental result show that the low-intensity particle recognition rate is improved from 79.8% to 97.7%.(3)The defect detection algorithm of TFT-LCD chips is studied.In view of the onchip defects and inter-electrode defects,a frequency domain filter is designed to eliminate the periodic stripe formed by the electrode,and gradient map is used to enhance the filtered image.The region growing method is used to segment the defect ROI region.The experimental results show that the edge of the ROI region obtained by this method is clearer than that obtained by threshold segmentation method.The application of SVM in defect classification is studied,and the feature-paralleled-OVRSVM is designed,which is trained separately for different types of features,and then the feature weight is used to merge the result of each SVM.The feature weight parameters are optimized on the validation set.The experimental results show that overfitting is avoided in the feature-paralleled-OVR-SVM,and the accuracy on the test set is improved from 81.7% to 96.8%.(4)The quantitative evaluation method of ACF interconnection quality is studied.A quantitative evaluation method is studied from three aspects: quantity,position and uniformity of conductive particles on the pad.In terms of uniformity,KL divergence between particle distribution and uniform distribution is used to evaluate the uniformity quality indirectly.Compared with Voronoi diagram method,the experimental results show that KL divergence method is more suitable for evaluating the uniformity quality.(5)The software platform of TFT-LCD chips detection is developed based on the laboratory environment.The functional requirements of the platform are analyzed and the overall architecture and technical configuration of the platform is designed.The platform is divided into five modules: data management module,task execution module,detection algorithm module,quality analysis module and real-time monitor module.The experimental results show that efficient and automatic detection can be realized by using the platform.