Anomaly detection poses significant challenges due to the scarcity and diversity of anomalous samples. Prior methods have demonstrated the efficacy of student-teacher networks in this domain. However, these approaches often suffer from false positives in noisy backgrounds due to inadequate feature discrimination. To overcome this limitation, we propose two key enhancements. First, we employ multi-feature fusion modules to extract and fuse low and high dimensional features, enabling superior reconstruction of low-dimensional textures and contours. Second, since synthetic anomaly images used as input can introduce feature redundancy detrimental to reconstruction, we incorporate a contrastive loss to encourage reconstructed features to resemble normal features while diverge from anomalous ones. On the MVTec-AD benchmark, our method attains remarkable results, including an image-level AUC of 99.2%, a pixel-level average precision (AP) of 79.3%, and an instance-level AP of 79.6%.