Development of Training-free Pattern Detection Methods for Manufacturing Inspection Images

Abstract

Manufacturing processes often need to check if a product meets specified shape requirements using inspection images. In this context, it is preferable to detect unknown shape patterns without additional learning when deploying machine learning models. However, existing object detection models struggle to identify patterns that were not previously learned. This study evaluated a template-based, training-free pattern detection method. We conducted experiments using 900 scanning electron microscope (SEM) images of semiconductor devices to assess methods that combine various algorithms, including a pre-trained deep learning model for few-shot object detection and template matching. The results indicated that the template matching method, featuring detection and removal components, achieved higher accuracy than deep learning models in SEM images. This difference is likely due to the deep learning models not including the specific pattern shapes found in SEM images during their training. In future work, we aim to explore enhancements in detection accuracy for deep learning-based methods and validate our approach using a broader dataset.