Development And Application Of Transfer Learning Based Intelligent Recognition System Of TFT-LCD Defect

In recent years, Thin Film Transistor-Liquid Crystal Display(TFT-LCD) has become the most mainstream of the display device, and toward high-resolution, large-size, light-oriented direction. The manufacturing process of TFT-LCD has become more and more complex, defect intelligent recognition system is becoming more and more important in the manufacturing to ensure product quality. In this paper, we designed a defect intelligent recognition system to meet the requirement of the testing process of TFT-LCD in Cell stage. The concrete research contents are as follows:On the basis of TFT-LCD defect testing requirements, we designed the overall scheme of the intelligent recognition system. According to the characteristics of the defect in Cell process stage, the hardware solution of the visual system was designed to ensure that the high-precision, high-contrast image acquisition and high-performance processing of the intelligent recognition system.According to the uneven phenomenon of the acquired image of visual system, optical model of each module of the hardware system were establish to analyze the factors which affect the image non-uniformity, and optimized these factors to improve the image uneven phenomenon.Traditional Mura defect classification algorithms can only identify few amount of Mura defect. And their recognition accuracy rate are not high enough. What's more, their training and recognition process are always time-consuming. According to above problems, we proposed a new kind of classification algorithm called online sequential classification based on transfer learning(OSC-TL) to achieve Mura defect online learning and recognition.According to the requirements of TFT-LCD defect inspection, the software system was designed and implemented, including software architecture design, the underlying interface design and human-computer interface design.The TFT-LCD defect intelligent recognition system have been tested in the actual production process. The defect recognition rate is 93%, and recognition time is about 30 milliseconds. The testing shown that the system has achieved the requirements of TFT-LCD defect online recognition.