Multi-scale Frequency-Decoupling and Prototype Alignment for Domain-Generalized Steel Surface Defect Detection

Steel surface defect detection is a critical task in industrial quality control. However, due to factors such as varying illumination, equipment differences, and diverse imaging modalities in real-world industrial settings, the generalization performance of traditional convolutional neural network (CNN)-based detection methods significantly degrades when confronted with domain shifts. Although existing domain adaptation methods can partially alleviate these issues, they require access to target domain data and exhibit limited adaptability to dynamic multi-domain environments. To address these challenges, this paper proposes a domain generalization detection model named MPADG, based on frequency domain decoupling and prototype alignment. Specifically, to tackle the problem of entangled defect features and domain-specific noise, a Multi-Scale Frequency Domain decoupling module (MSFD) module is designed to separate defect-relevant features from domain-related noise. To resolve the issues of scattered cross-domain feature distribution and blurred boundaries, a Prototype-Aligned Feature Aggregation Module (PAFAM) is employed to enhance cross-domain semantic consistency. Furthermore, an Adaptive Prototypical Contrastive Loss (APCL) is introduced to mitigate the class imbalance problem. Experiments conducted on the NEU-DET dataset and its extended versions incorporating underwater and foggy domain data demonstrate that MPADG achieves an average detection accuracy of 72.61% across multiple unseen target domains. This result significantly outperforms existing domain generalization methods, validating its superior generalization capability and robustness.