A Unified Wafer Anomaly Detection and Segmentation

We present UniWADS, a unified framework for wafer anomaly detection and segmentation in semiconductor high-volume manufacturing (HVM). Existing automated optical inspection (AOI) pipelines rely on hand-crafted rules, defect-specific thresholds, or class-specific training, limiting scalability across products and process conditions. UniWADS addresses this by performing image-level anomaly detection and pixel-level localization within a single reconstruction-based architecture.UniWADS follows a constrained reconstruction principle: robust anomaly detection requires both stable token correspondence and controlled decoder capacity. It combines a frozen EfficientNet-B4 encoder, a lightweight SpatialAdapter decoder with full softmax attention, a zero-parameter Noisy Bottleneck, and a hard cosine reconstruction objective to suppress identity mapping and amplify defect-sensitive discrepancies.We evaluate UniWADS on Texture-AD Wafer and MIIC/IC microscopy datasets. Ablations show that softmax attention is more reliable than linear and hybrid variants, while the Noisy Bottleneck and hard cosine loss play complementary roles: the former improves localization stability, and the latter enhances image-level separability. On Texture-AD Wafer, the best configuration achieves 82.46% pixel-AUROC and 58.36% AUPRO.These results demonstrate that constrained reconstruction provides a robust and deployment-friendly solution for unified wafer anomaly inspection. More fundamentally, we show that reconstruction-based anomaly detection is governed by a capacity–correspondence trade-off: limiting reconstruction capacity while preserving token correspondence is key to separating normal and anomalous patterns. The code is available at https://github.com/leolin65/UniWADS.git.