Answers are grounded in this paper and cite their sources.
Thin film transistor liquid crystal display (TFT-LCD) manufacturing is capital intensive with complex high-tech production processes. To maintain competitiveness, conventional vertical-integrated companies have been replaced by collaborations among many complementary and specialized companies in the ecosystem. TFT-LCD manufacturers strive to enhance advanced quality control via accurately detecting various defects throughout production for yield enhancement and cost effectiveness. Focusing on realistic needs for TFT-LCD manufacturing, this study aims to develop a unified framework for digital decision support for known defect classification and unknown defect rejection with limited and imbalanced training data in real settings. In particular, the training phase includes two main steps. Firstly, deep transfer learning is developed for training discriminative models with known defect data. Secondly, convolution prototype networks will be developed where the model weights can be initialized by the trained discriminative model in step 1 with different distance-based loss functions. In the testing phase, unknown defects can be identified via prototype networks based on distance-based rejection threshold. If test samples are classified as known defects, final predictions will be derived via ensemble learning of multiple classifiers with the features extracted from both discriminative and prototype model. If the samples are uncertain to reject as unknown defects, a novel two-stage rejection strategy is proposed to enhance the unknown defect rejection accuracy. To validate the proposed framework, an empirical study is conducted in a leading TFT-LCD manufacturer in Taiwan that has shown significant improvements than the existing approaches for advanced quality control.
No takes yet. Share an insight, caveat, or question.
Trang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: