Study on Anomaly Classifier with Domain Adaptation

There are various characteristics of industrial defects, and there is no fixed pattern to search for. Typically, anomaly detection models are used to identify defects. However, after experiencing a domain gap, industrial defect images often lead to a decrease in the verification accuracy of the source model. We conducted experiments to validate this and employed a domain adaptation model. Using color transformation algorithms, we generated source images with domain gaps and introduced them to a pre-trained model. We trained the model to learn features from the source domain and utilized a domain discriminator to differentiate between features from the source and target domains, assuming that the mappings of the target and source domains come from the same distribution. Comparative experimental results demonstrate that the domain adaptation model has a significant impact on improving accuracy. Specifically, the accuracy of the original “flower” category increased from 43.98% to 89.23%, and the “cable” category improved from 75.33% to 85.66%.