Anindita Suryarasmi

Defect classification for thin-film transistor liquid crystal displays (TFT-LCD) poses significant challenges due to the fine-grained nature of microscopic defects and the limited availability of labeled data. This problem is particularly demanding in the context of few-shot class-incremental learning (FSCIL). FSCIL requires learning new defect classes from limited data while distinguishing between highly similar classes. At the same time, it must preserve knowledge of previously learned classes. We introduce Fine-grained Few-shot Class Incremental Learning with Destruction-construction Learning (FFCIL) to address these challenges. This approach employs a multistage training strategy that begins with destruction-construction learning. This stage captures intricate and fine-grained features during pretraining and is combined with manifold-mixup as a placeholder mechanism. Together, these components enable the seamless integration of new defect classes with minimal samples. Furthermore, a triplet loss is applied during pseudo-incremental learning to improve the model’s ability to differentiate between base and pseudo-classes. This approach enhances the fine-grained classification performance in a few-shot class incremental learning scenarios. Experimental results on an industry-based TFT-LCD dataset demonstrate a notable 16.36% improvement in overall accuracy. In addition, FFCIL achieves the lowest performance dropping rate among all compared methods, indicating superior stability. Tests on public datasets further reveal an improvement of up to 39.58% in overall accuracy, underscoring the potential applicability of FFCIL to broader domains. Code is available at https://github.com/sallylinlab/ffcil