The Research Of The Defects Detection Technology Of The TFT-LCD

Thin Film Transistor-Liquid Crystal Display (TFT-LCD for short) is famous for its high resolution, high illumination, and non-geometry-distortion. And because of its small bulk, light weight and low power consumption, TFT-LCD now is been widely used in the display field. The processing of making TFT-LCD which includes more 100 working procedures is very complicated. Therefore, it's hard to avoid that there are some defects on the panels while making the TFT-LCD. The detection of the defects on the TFT-LCD panels is very helpful for the statistic analysis, fault repair and washing out the rejects.Pinhole, scratch and particle are familiar defects of the TFT-LCD. While taking photos for the TFT-LCD, the images turned out to be uneven brightness and aperiodic which was caused by the physical trait of the camera. Furthermore, the high resolution of the camera led to the large size of each image which was a disadvantage of the real time processing. Therefore, uneven brightness and aperiodic of the image, the efficiency of the system are the main problems to be solved in this article.In this article, we designed a defects detection system based on the machine vision. In allusion to the difficulty of uneven brightness, we parted the image into areas by the algorithm based on Average Sampling to overcome it firstly. For solving the problem of aperiodic, we segmented the image using the algorithm based on Projection Valley and classify the regions. Finally, we detected the defects by extracting the features and using the BP neural network.Aiming at the disadvantage of the physical trait from the camera, we shorten the distance of the TFT-LCD panel and the camera to reduce the influence of the uneven brightness and the aperiodic of the image, firstly.For overcoming the influence of the uneven brightness, in this article, we analyze the image macroscopically and proposed an algorithm to part the image which was based on Average Sampling. The arithmetic used a point to represent a region of the original image. The gray level about that point was the average gray level of the region. Through this algorithm, we not only parted the image effectively by setting the threshold, but also enhanced the efficiency of the system. After parted the image based on the Average Sampling, we gained the areas that the brightness were even, which could overcome the disadvantage of uneven brightness effectively.Aiming at reducing the impact of the aperiodic, we used another algorithm to segment the areas, which based on the Projection Valley, by analyzing the areas microscopically. By projecting the areas with the Row and Column and getting the projection curves, the arithmetic got the valleys of each curve. After that, the arithmetic segmented the areas into lots of regions by the valley points, based on the features of the areas inside. Finally, we classify the regions by their structure so as to overcome the problem of aperiodic.After analyzing the regions, we got the features of the regions. With the theory of the image features, we got the area and the structures of each region as the feature vector for the defects detection.Because the illumination of each image and the area after segmenting is different, it is very difficult to set the threshold for designing the detection arithmetic. As the BP neural network has the characters of self-organization, self-adaptation and self-learning, we designed a BP neural network to train the vectors which were normalized. After that, we recognized the region defects by the parameters of the neural network. Finally, we got the exactly position of the defects from the regions by comparing the non-defect region with the defect region.In this article, we analyzed the image macroscopically and microscopically and proposed the algorithm of image segment based on Average Sampling and Projection Valley. It overcame the disadvantage of the uneven brightness and the aperiodic effectively. Finally, we detected the defects by features extracting and the neural network. Experiences showed that the detecting system we designed could detect the defects quickly and effectively.