Selective Auto-Labeling for Deep Learning-Based Domain-Specific Image Classification in Manufacturing

In the realm of semiconductor process control, the rapid and accurate identification of issues through historical equipment maintenance imagery is crucial for timely corrective actions. However, the application of Convolutional Neural Networks (CNNs) for image classification within this domain is hampered by the scarcity of labeled data, a consequence of the highly specialized and complex nature of semiconductor manufacturing processes. This paper introduces a novel approach to mitigate the challenge of minimal labeling by implementing a confidence-based auto-labeler. Leveraging domain-specific knowledge and a minimal set of 13 labeled images, we successfully generate a sufficiently large and reliable labeled dataset. This dataset serves as the foundation for training CNN-based models, demonstrating the potential for significant improvements in decision-making efficiency in semiconductor process control. Our methodology not only addresses the labeling bottleneck but also sets a precedent for employing machine learning and AI in enhancing the robustness of semiconductor manufacturing and maintenance practices.