近年來，隨著深度學習技術的蓬勃發展，無監督異常檢測已逐漸應用於工業檢測 與醫學影像等高精度需求的場景，特別是在缺乏大量標註資料的情況下，成為降低 標註成本的重要方向。傳統的重建式方法透過將輸入還原為重建圖以偵測異常，雖 在簡單資料集上具備一定成效，但面對真實異常時，仍經常出現語意坍塌（Semantic Collapse）問題，導致模型無法準確區分正常與異常區域。 為克服上述挑戰，本研究提出一種結合合成異常與真實異常訓練策略的US模組 （Unsupervised Segmentation Module），搭配多層特徵融合架構進行異常分割。該方法以 預訓練特徵為基礎，搭配重建型分割模組進行監督學習。此外，為避免重建模組產生 語意坍塌現象，本研究結合合成異常擾動與特徵對齊策略，促使模型在保持正常樣本 還原品質的同時，也能辨識微小異常區域。 在實驗設計上，我們以非公開產線面板資料集A19作為主測試資料，並輔以Retina 醫學影像資料集、MVTec-Leather與MVTec-Bottle兩種工業資料集進行驗證。實驗結果 顯示，本研究方法在大多數資料集中皆可達到最佳或次佳之pixel-AUROC與AP指標， 顯示本方法在無標記異常下具備強健的學習能力。與主流方法比較，本研究方法在異 常定位上更具辨識力。 綜合上述結果，本研究提出之US模組能有效緩解語意坍塌問題，並於不同資料型 態與異常類型中展現穩定且優異的異常分割表現，具備應用於實際工業與醫療場域之 潛力。;In recent years, the rapid advancement of deep learning has led to significant progress in unsupervised anomaly detection, especially in fields such as industrial inspection and medical imaging where labeled data is often scarce or expensive to obtain. Traditional reconstruction based approaches rely on reconstructing input images to identify anomalies. However, these methods often suffer from semantic collapse, resulting in poor detection and localization per formance. To address this issue, we propose a novel Unsupervised Segmentation (US) module that combines synthetic and real anomalies in a unified training scheme. The method leverages pretrained visual features and integrates a reconstruction-based segmentation decoder with a multi-layer feature aggregation design. Furthermore, our method employs anomaly augmenta tion and feature alignment to suppress the semantic collapse effect and enhance the model’s ability to localize fine-grained anomaly regions. We validate our method on multiple datasets, including a private industrial panel dataset (A19), a medical Retina dataset, and two representative MVTec datasets (Leather and Bottle). Experimental results demonstrate that our method consistently achieves competitive or superior performance in terms of pixel-level AUROC and AP. Compared to existing approaches, the proposed method excels in segmentation accuracy and robustness. In conclusion, the proposed US module effectively alleviates the issue of semantic col lapse in reconstruction-based methods and demonstrates strong and stable performance across diverse anomaly types and real-world datasets, indicating its potential for practical deployment in industrial and medical anomaly detection tasks.
