High-precision visual inspection of semiconductor optical devices is critical but challenged by the coexistence of local and global anomalies, alongside domain-specific imaging artifacts. This study proposes a frequency-specialized dual-branch anomaly detection framework integrated with domain-specific preprocessing to address these issues. First, a Masked Autoencoder (MAE)-based preprocessing module reconstructs peripheral shadow artifacts to eliminate boundary-induced false positives. Then, utilizing the Stationary Wavelet Transform (SWT), the method explicitly decouples defect signatures into frequency domains. Low-frequency components are routed to an MAE-based global branch, while high-frequency components are processed by a Reverse Distillation (RD++)-based local branch. Experimental results on an industrial dataset demonstrate that the proposed method achieves an AUROC of 0.9837 and an AP of 0.9701, significantly outperforming the baselines. By attaining a 100% detection rate for well-defined local defects, the framework demonstrates its high reliability for strict industrial quality control where minimizing false negatives is paramount.