Anomaly detection in computer vision is typically framed as a one-class classification and segmentation problem. While reconstruction-based methods are widely used, existing networks often suffer from generalization on test data and sensitivity to background noise. We propose a two-stage framework with encoder freezing and hybrid learning, which comprises three key components: (1) a pre-trained encoder freezing to guide normal feature learning, (2) odd-even dataset partitioning for separate reconstruction and discrimination training, and (3) hybrid inputs to enhance noise robustness. Evaluation on industrial inspection benchmarks shows that our method achieves outstanding performance with 98.2% image-level AUC and 75.1% pixel-level AP.