The detection of surface defects in casting materials presents significant challenges due to coarse textures, irregular patterns, and class imbalance, which hinder the development of highly reliable automatic classification methods. To address these challenges, this paper proposes a dual-encoder-based generative anomaly detection framework using only normal data. Specifically, the proposed method is designed to integrate pixel responses, critic intermediate features, and latent representations into a unified anomaly score and to enforce consistency across input-reconstruction-latent spaces. A dual-encoder architecture combined with a Wasserstein GAN with gradient penalty backbone is proposed to effectively learn and align these multi-perspective representations. Both quantitative and qualitative results demonstrate that the proposed model achieves superior performance in defect classification and reliability, achieving an Area under the receiver operating characteristic curve of 0.998 and an accuracy of 99.1%. Balanced values in precision, recall, and F1-score indicates effective suppression of false positives and prevention of missed detections, showing clear improvements compared to baseline models. This study not only enables precise detection of surface defects in casting processes but also holds potential to contribute to anomaly detection across diverse manufacturing processes.
Kwak et al. (Fri,) studied this question.