A Comparative Study of Deep-Learning Object Detectors for Semiconductor Defect Detection

Deep learning has become the de facto method for natural image object detection. Recent work has also shown that deep-learning object detectors can outperform traditional image processing methods for semiconductor defect inspection. In this study, state-of-the-art (SOTA) deep-learning object detection models were trained on a dataset of images with line space pattern defects towards improving average precision (AP). Experimental results show that DINO performs best with a mean AP of 0.865 with an inference time of 108.7 ms/image, while YOLOv7 achieves a good balance between precision (0.843 mAP) and inference time (20.2 ms/image). Furthermore, the results obtained from this study suggest that focusing on small, challenging classes and ensembling different model types can improve performance.