Automatic target defect identification for TFT-LCD array process inspection using kernel FCM-based fuzzy SVDD ensemble

TLDR: Experimental results indicate that the proposed F-SVDD ensemble not only greatly improves the performance of SVDD, but also outperforms other commonly used classifiers such as support vector machine (SVM), in terms of target defect identification rate.

Abstract: Inline defect inspection plays a critical role in yield improvement for thin film transistor liquid crystal display (TFT-LCD) manufacturing. In array process, some defects are critical to the quality of LCD panels (target defects), while some are not (non-target defects). This paper proposes a target defect identification system by which the target defects can be automatically identified. The proposed system is composed of five parts: projection-based pixel segmentation, normal pixel removal, feature extraction, target defect identification, and decision making. For the identifier design, a novel one-class kernel classifier called fuzzy support vector data description (F-SVDD) ensemble is proposed. F-SVDD ensemble is proposed to solve two critical problems existing in SVDD, including the overfitting due to outliers, and the multi-cluster distribution. In F-SVDD ensemble, both the best number of the F-SVDD members in the ensemble and the elements of each member can be determined by using partitioning-entropy-based kernel fuzzy c-means (KFCM) algorithm. Experimental results, carried out by real defective images provided by a LCD manufacturer, indicate that the proposed F-SVDD ensemble not only greatly improves the performance of SVDD, but also outperforms other commonly used classifiers such as support vector machine (SVM), in terms of target defect identification rate. In addition, the task of target defect identification for one defective image can be accomplished within 3s by the proposed system.