CRASDG: Clustering and Reconstruction-based Adversarial Single Domain Generalization Network for Die Casting Defect Detection

Defect detection in the die casting production process plays an important role in enhancing factory efficiency. Existing deep learning-based detection methods rely on sufficient annotated data for all types of products, while a common case is that annotated samples are only available for a specific type of product, which can be regarded as a single domain generalization (SDG) problem. To address this issue, this work proposes a novel clustering and reconstruction-based adversarial single domain generalization (CRASDG) network for die sasting defect detection. In CRASDG, a clustering process is first conducted on the statistics of layers of the feature extractor to obtain pseudo domain labels. Then, a reconstruction process maximizes the distance between reconstructed samples belonging to different pseudo latent domains. In addition, the adversarial training strategy helps the feature extractor learn domain-invariant features. Experimental results in a real die-casting process demonstrated the superiority of the proposed CRASDG method over other existing methods.