Unsupervised anomaly detection methods based on knowledge distillation have demonstrated significant advantages in industrial quality inspection, medical diagnosis, and related fields. However, due to the lack of discriminative abnormal features in the student network and the loss of semantic information during feature reconstruction, detection accuracy remains limited. To address this issue, a semantic-guided denoising knowledge distillation model for anomaly detection is proposed. Firstly, an anomaly synthesis algorithm based on Perlin noise is introduced to provide abnormal discriminative cues for the denoising student network, thus enhancing the feature discrepancy between teacher and student networks and mitigating overfitting. In addition, a semantic-guided network with spatial-aware feature fusion module is designed to connect the denoising student network, thus preserving semantic integrity and improving fine-grained reconstruction capability of anomalous regions. Finally, a lightweight adaptive segmentation network is developed based on multi-scale features from both teacher and student networks to enable accurate anomaly localization. The experimental results show that the proposed method achieves an pixel-level AP of 78.0% on the MVTec AD dataset and 44.1% on the VISA dataset, outperforming all existing advanced methods.