A domain generalization classification algorithm for surface defect of photovoltaic silicon wafer based on mixup and DIFEX

In the surface defect detection of photovoltaic silicon wafers, detection targets exhibit significant morphological variations. While deep learning-based classification models outperform traditional methods in accuracy, the inherent domain shift caused by divergent image distributions across domains substantially degrades model performance in cross-domain scenarios. To enhance model generalization capability, a domain generalization classification method based on Mixup and DIFEX (Domain-Invariant Feature EXploration) is proposed. Firstly, a dual distillation strategy is used to derive both internally invariant features (image phase and low-frequency information) and mutually invariant features (image Mixup fusion information). Subsequently, the regularization term is introduced to maximize the discrepancy between these two feature representations. Finally, the overall training process is completed by minimizing the composite loss function combining feature discrepancy regularization and classification error. Experimental results on both the self-made photovoltaic wafer defect dataset and the PACS benchmark demonstrate three key findings: (1) The proposed method achieves consistent Top-1 accuracy improvements of 1.05% and 9.81% over conventional approaches on respective datasets; (2) Compared with standalone Mixup implementation, it shows accuracy gains of 0.16% and 4.15% across testing domains; (3) When benchmarked against pure DIFEX framework, the hybrid approach exhibits superior performance with 1.39% and 2.53% enhancement in generalization. These statistically validated improvements demonstrate that the integrated methodology achieves competitive generalization performance, advancing the current capabilities in surface defect classification tasks.