This paper aims to solve the problem of unsupervised representation learning for wafer defect recognition.
The authpors propose a defect pattern recognition method based on invariance propagation (DPRIP). This method consists of an unsupervised pre-training model and a fine-tuning model. In the pre-training stage, the authors make full use of readily available unlabeled data to learn rich feature representations through the invariance propagation algorithm, and then they serve as initialization of convolutional neural network classification model. After unsupervised pre-training, the authors use a small fraction of labeled data to fine-tune the classification model to get the final network model.
The authors apply DPRIP to semiconductor defect images and test it on the real-world publicly accessible wafer map dataset WM-811K. The experimental results outperform the current state-of-the-art methods, which indicates the effectiveness and applicability of this method.
The authors propose a new surface defect pattern recognition model based on unsupervised representation learning, which can significantly improve the performance of the recognition task when labels are scarce. The authors analyze the role of temperature in unsupervised training and find that it is the key parameter to control the attention of the model to the hard negative sample images. An appropriate temperature can better separate target defect pattern from other similar defect patterns. The authors use unlabeled data to learn rich feature representations in advance, which not only significantly reduces the impact of human priori but also effectively solves the high cost problem of manual supervision. The authors validate the effectiveness of our method through quantitative experiments, and they achieved state-of-the-art results on the public wafer map data set (WM-811K).
