<p>Current mainstream unsupervised anomaly detection methods typically extract raw features using pre-trained networks. However, overlooking domain bias and underutilizing extracted features hinder detection accuracy. To address overfitting from single-class learning, many approaches employ defect synthesis strategies to improve detection. Yet, existing methods depend significantly on abnormal textures from external datasets. Additionally, limited coverage and directionality in synthesis can miss key features and introduce redundancy. To address these challenges, we propose a novel unsupervised framework, MSAF-SWCS, that jointly leverages attention-based feature refinement and dual-level synthetic anomaly generation to tackle industrial defect detection. A core innovation lies in the Multi-Scale Attention Fusion (MSAF) module, which adaptively enhances spatial and channel representations across hierarchical features. Another key novelty is the Strong and Weak Defect Collaborative Synthesis (SWDCS), which unifies image-level and feature-level anomaly simulation to mimic diverse industrial defects more realistically. MSAF-SWCS reaches 99.5% AUROC at the image level and 94.1% AUPRO at the pixel level on the MVTec AD dataset. On the VisA dataset, it attains 98.2% AUROC and 94.2% AUPRO, respectively. Furthermore, the proposed method demonstrates strong generalization ability on more realistic industrial benchmarks, achieving an average image-level AUROC of 91.9% and pixel-level AUPRO of 90.4% on the Real-IAD dataset, as well as 94.5% AUROC and 78.9% AUPRO on the BTAD dataset. The results validate MSAF-SWCS as a practical and effective solution for industrial scenarios.</p>