Research Of Defect Inspection Of TFT-LCD Based On MapReduce

The size of the liquid crystal display is getting bigger and bigger with the continuous development of production technology of TFT-LCD. The sixth generation of TFT-LCD with the latest process technology can reach a size of1500mm×1800mm. As a result, the accuracy and speed of traditional automatic defect inspection method with one single CCD camera can hardly satisfy the requirements of the current and future inspections. To solve the problems of large field of view and high inspection precision in the inspection of big size LCD, one possible solution is to use multiple CCD instead of a single CCD to scan the LCD images. To be specific, multiple linear array CCD are lined up in a row to scan the LCD for one or more times in one inspection. However, the multiple split images captured by the CCD is a challenge to the limited processing capacity of a single computer or an embedded processor. Hadoop, with huge storage capacity and parallel computing capability, is a feasible new way to deal with massive high-resolution LCD images.The goal of this thesis is to improve the automatic defect inspection process of the LCD image based on the Hadoop platform. The advantages of the storage capability and distributed processing technology of Hadoop can be utilized to improve the speed and accuracy of defect inspection. Distributed defect inspection of LCD images was combined with Hadoop in this thesis, the main work is as follows:(1)A deep research and analysis of the Hadoop platform, especially on the two core parts of the platform:HDFS distributed file system and MapReduce parallel programming framework, has been conducted. It laid the theoretical foundation and technical support for the distributed storage and processing of the LCD image.(2)In view of the defect inspection process, we made an intensive study of these four processes:image background texture suppression, defect segmentation, feature extraction and classification. Texture suppression algorithm based on MapReduce has been implemented. And we designed a new file storage format for massive images based on which the addressing and reading speed of image processing on MapReduce would be improved.(3)The segmentation method of TFT-LCD with blurry contour had been mainly studied. Different from the traditional C-V segmentation model, we added constraint term and the gray level difference to obtain a better segmentation result. Afterwards, the second segmentation had been executed for the defect at the edge of the sub image by using MapReduce framework. Finally, we got the entire defect region. After extracting feature from the defect region, we used SVM classification based on single-class method to classify defects.The experiments demonstrated that this method could inspect defects with a good speed up rate and reduce the miscarriage of justice for defects simultaneously.