Abstract:

Unsupervised methods for defect detection hold significant potential for automated optical inspection (AOI) due to their independence from the challenges of collecting defect data. However, their performance is often limited in scenarios where normal samples exhibit high variability and defect features are subtle. To address this limitation, an unsupervised defect detection method named feature isolation autoencoder (FIAE) is proposed. This method compresses images into two components: one that encodes the information necessary to reconstruct normal samples, and the other that encodes anomalous information. As a result, any non-zero value in the anomalous component indicates the presence of an anomaly in the input images. Additionally, we propose the dynamic defect variability guidance (DDVG) technique and the stabilization-compensation training strategy, which ensure that FIAE achieves the expected training outcomes in the training data and process levels, respectively. Extensive experiments on the transistor dataset compiled in this study and the benchmark dataset MVTec Anomaly Detection dataset demonstrate that FIAE outperforms the baseline methods in most cases, achieving an area under the curve (AUC) of 0.979 and a best F1 score of 0.958 on the transistor dataset. The necessity of the proposed DDVG and the stabilization-compensation training strategy through ablation studies are validated.

Key words: defect detection, unsupervised learning, convolutional autoencoder