SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection

2026-07-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

SwinAD is a novel reconstruction-based framework designed for multi-class unsupervised industrial anomaly detection, addressing challenges where diverse normal patterns can lead to over-generalization. This system employs a frozen pretrained Swin Transformer V2 encoder to extract semantically rich multi-scale features. It pairs this with a feature diversity-preserving reconstruction decoder, which includes stage-wise bottleneck modules with dropout to prevent trivial identity mapping and ensure robust reconstruction of normal patterns. A key innovation is its ability to maintain complementary reconstruction hypotheses, moving beyond single decoding branches to improve localization. The final anomaly map is generated by aggregating discrepancies between the encoder features and two reconstructed features across multiple scales. Evaluated on MVTec AD, VisA, and Real-IAD benchmarks, SwinAD demonstrates competitive image-level performance and strong pixel-level localization accuracy, particularly improving pixel-level AP and 1 on MVTec AD.

Key takeaway

For Machine Learning Engineers developing industrial inspection systems, SwinAD offers a robust approach to enhance anomaly detection, especially in multi-class scenarios. If your current methods struggle with diverse normal patterns or pixel-level localization, consider integrating multi-stage feature reconstruction with a frozen Swin Transformer V2 encoder. This can significantly improve pixel-level accuracy, as demonstrated by SwinAD's performance on benchmarks like MVTec AD.

Key insights

SwinAD uses a Swin Transformer V2 encoder and diverse multi-scale reconstruction to improve pixel-level anomaly localization in multi-class industrial settings.

Principles

• Diverse reconstruction improves pixel-level localization.
• Multi-scale features enhance anomaly detection.
• Bottleneck modules prevent trivial identity mapping.

Method

SwinAD reconstructs multi-scale features from a frozen Swin Transformer V2 encoder using a diversity-preserving decoder. Discrepancies between original and two reconstructed features are aggregated across scales to form an anomaly map.

Topics

• Anomaly Detection
• Industrial Inspection
• Swin Transformer V2
• Feature Reconstruction
• Multi-class Learning
• Pixel-level Localization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Statistical Non-linear Reconstruction Loss for Image Anomaly Detection — Non-linear reconstruction loss with statistical calibration prevents outlier leakage in image anomaly detection.
• DifferAD-R1: A Difference-Guided IndustrialAnomaly Localization with Multimodal LargeLanguage Models — DifferAD-R1 uses a difference-guided MLLM-augmented RL framework to localize industrial anomalies, improving generalization and detection of subtle defects.
• RAD: A Dataset and Benchmark for Real-Life Anomaly Detection with Robotic Observations — RAD dataset bridges the gap between idealized and real-world industrial anomaly detection by incorporating robotic observation noise.
• Rethinking Infrastructure Inspection as Image Difference Classification: A Traffic Sign Case Study — Image difference classification with instruction-based models reduces data dependency for infrastructure defect detection.
• Visual Prompting Meets Feature Reconstruction-Based Anomaly Detection with Dual-Teacher Supervision — Robust anomaly detection in varied real-world conditions is achieved by combining visual prompting, adaptive teacher models, and synthetic data.
• OPD-IAD: From Language Judgment to Industrial Anomaly Detection via On-Policy Self-Distillation — OPD-IAD precisely localizes industrial anomalies by distilling language judgments onto visual features via on-policy self-distillation and visual anchoring.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.