Automatic Labeling in Image Segmentation and Classification for TFT-LCD Manufacturing

Abstract

Product quality inspection of Thin-film transistor-liquid crystal display (TFT-LCD) is time-consuming and labor-intensive. An automatic algorithm-based defect inspection system solves these problems by reducing the time and manual labor involved. This research work proposes an AI-based LCD inspection vision system that evaluates defects by AI-based defect segmentation and classification. Time-consuming pixel-level labeling of defects can be eliminated by applying weakly-supervised learning methods. The efforts of quality control personnel for model training can be reduced which is especially important in high-mix production with a high amount of changeovers and production process adjustments. Therefore, PP-CAM (Precise-Puzzle-CAM) based defect segmentation method is proposed to cope with diverse TFT-LCD defect shapes and sizes. Secondly, a region of interest-supported classification method is developed to enable cropping of small-scale TFT-LCD. The performance of the AI methods are investigated using industrial TFT-LCD manufacturing datasets of two manufacturing processes.