Auto-Labeling of IC Markings via Reinforcement Learning

To achieve high accuracies in deep learning-based inspection of IC markings, a large number of manually labeled images of good and defective markings are needed. In many IC production lines, only a handful of labeled images of different defect types are available that are manually labeled. This paper presents an auto-labeling method in order to assign labels to a large number of unlabeled images in an automatic manner. Interpretable defect-specific rules incorporating threshold values are considered to identify different defect types. These rules are applied in a hierarchical manner to detect defective markings as well as good markings. A reinforcement learning framework is developed to obtain the optimum values of the thresholds based on the safe-bandit update mechanism. Experimental results on a blind study dataset demonstrate that the developed auto-labeling method, named ARL, achieves an all-class and a binary labeling accuracy of 92.1% and 94.7%, respectively. This and other results show that ARL provides a practical and interpretable solution for automating the labeling of IC markings.