Feature Clustering for Open-Set Recognition in LCD Manufacturing

Inspecting defects in liquid-crystal display (LCD) manufacturing is of uttermost importance to ensure customer's satisfaction and reduce time and money losses. Deep learning classification methods rely on the closed-set assumption that the classes to predict during operation are the same as the training ones. However, in real-world settings, new unseen classes (defects) often arise. In this work, we evaluate the capabilities of state-of-the-art deep learning methods for classifying known and unknown defects on LCD images. Given the limited performance of such methods, we here propose a novel cluster error (CE) classifier and a strong-repulsive (SR) training loss for feature clustering to enhance the classification accuracy both on known and unknown defects. Our results on two real-world industrial datasets show the challenges of such task and how our classifier outperforms the other methods.