💻 computer science

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

SwinAD is a multi-stage feature reconstruction framework that leverages a frozen Swin Transformer V2 encoder and a diversity-preserving decoder to achieve competitive image-level detection and superior pixel-level localization in multi-class unsupervised industrial anomaly detection.

Original authors: Huong Ninh, Chien Thai, Mai Xuan Trang, Vu-Minh Le, Thanh Ha Le, Long Tran

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Huong Ninh, Chien Thai, Mai Xuan Trang, Vu-Minh Le, Thanh Ha Le, Long Tran

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the head of quality control at a massive, high-tech factory. Your job is to spot the "bad apples" in a sea of perfect products. In the old days, you'd need a team of experts to look at every single item and draw a circle around every scratch, dent, or crack. But here's the problem: defects are rare, they come in thousands of weird shapes, and you can't possibly train a human (or a computer) to recognize every single type of mistake before it happens. So, scientists came up with a clever trick: instead of teaching the computer what a "broken" thing looks like, they only show it a million examples of "perfect" things. The computer learns the pattern of perfection so well that when it sees something that doesn't fit, it screams, "Hey, that's weird!" This is called unsupervised anomaly detection. It's like teaching a security guard to recognize the sound of a normal factory floor; if a weird noise pops up, the guard knows something is wrong without ever having heard that specific noise before.

But there's a catch. Most of these "perfect pattern" computers are like specialists who only know one type of product. If you switch from making glass bottles to making metal nuts, you have to build a whole new computer from scratch. That's slow, expensive, and a headache for big factories that make dozens of different things. Researchers are trying to build one "super-brain" that can handle all the different products at once. The challenge is that when you mix different products together, the computer gets confused. It might think a normal pattern from one product is a defect in another, or it might get so general that it misses tiny, subtle scratches. The big question is: How do we build a single brain that is smart enough to know the difference between a "normal" metal nut and a "broken" glass bottle, while still being sharp enough to spot a hairline crack?

Enter SwinAD, a new method proposed by researchers Huong Ninh, Chien Thai, and their team. Think of SwinAD as a high-tech detective that uses a special kind of "super-vision" to solve the mystery of industrial defects. Instead of trying to memorize every single product, it uses a pre-trained "brain" (called a Swin Transformer) that has already learned how to see the world in layers, from tiny details like texture and edges to big-picture shapes and meanings. This brain is frozen, meaning it doesn't learn new things during the test; it just provides a stable, reliable view of the image.

The real magic of SwinAD happens in how it tries to "reconstruct" or rebuild the image it sees. Imagine you are looking at a perfect puzzle piece. A normal computer might try to copy it exactly. But SwinAD is smarter. It uses a special trick called feature diversity-preserving reconstruction. Instead of trying to build just one copy of the perfect piece, it builds two slightly different copies at the same time. It's like asking two different artists to draw the same perfect apple. One might focus on the shine, the other on the curve. As long as both drawings look like a valid apple, the computer knows the original was normal. But if the original piece was actually a potato with a scratch, neither artist can draw a perfect apple from it. The mismatch between what the computer sees and what it tries to draw becomes a bright red flag.

By comparing the original image against these two different "hypotheses" of what a normal object should look like, SwinAD can spot even the tiniest, most subtle defects. It doesn't just say "This image is bad"; it draws a precise map showing exactly where the scratch or crack is. The researchers tested this on three major industrial datasets (MVTec AD, VisA, and Real-IAD), which are like the "Olympics" of defect detection. They found that SwinAD is incredibly good at pinpointing defects, often beating the current best methods. Even more impressively, it does this while using less computer power and running on smaller image sizes than its competitors. It suggests that by keeping the "super-vision" brain frozen and using a dual-drawing strategy to preserve variety, we can build a single, efficient system that keeps our factories running smoothly, spotting the bad apples without needing a separate expert for every single fruit in the basket.

Technical Summary: SwinAD for Multi-class Unsupervised Industrial Anomaly Detection

Problem Statement
Industrial anomaly detection (IAD) aims to identify and localize defective regions without relying on exhaustive annotations of all possible defect types. While recent unsupervised methods have achieved strong performance, most are designed for Single-class Unsupervised Anomaly Detection (SUAD), requiring a separate model for each object category. This approach incurs high computational and memory costs when scaling to industrial systems with numerous product types. Consequently, Multi-class Unsupervised Anomaly Detection (MUAD), which employs a single unified model across multiple categories, is preferred but remains challenging. The diversity of normal patterns across different classes can lead to over-generalization, reducing the discriminative capability between normal and anomalous regions. Furthermore, existing reconstruction-based methods often prioritize high-level semantic representations for global discrimination, which may suppress the fine-grained spatial details necessary for precise pixel-level localization.

Methodology: SwinAD
The authors propose SwinAD, a hierarchical reconstruction-based framework designed for multi-class unsupervised anomaly detection. The framework consists of three primary components:

Hierarchical Feature Extraction:
SwinAD utilizes a frozen, pretrained Swin Transformer V2 as the image encoder. Unlike methods using standard Vision Transformers (ViT) that focus on global semantics, the Swin Transformer's hierarchical structure provides multi-scale features ( $F_1$ to $F_4$ ) that capture both local textures (in shallow layers) and semantic structures (in deeper layers). To enhance robustness, the authors aggregate features across all blocks in the third stage rather than relying solely on the final block output. The encoder remains frozen during training to preserve stable, discriminative representations learned from large-scale data.
Feature Diversity-Preserving Reconstruction:
To address the issue of deterministic feature collapse—where a single decoder learns to reconstruct all normal patterns into a single prototype, failing to capture the manifold diversity of normal data—SwinAD employs a dual-decoder architecture. Instead of a single reconstruction branch, the framework maintains two complementary reconstruction hypotheses ( $D_1$ and $D_2$ ).
- Bottleneck Modules: Stage-wise bottleneck modules with dropout are inserted between the encoder and decoders. The dropout prevents trivial identity mapping, forcing the decoders to learn robust reconstructions of normal patterns rather than simply copying encoder features.
- Diversity Regularization: A cosine-based divergence loss is introduced to explicitly discourage the two reconstructed features from collapsing into identical representations. This ensures the model preserves a broader manifold structure of normal patterns, improving generalization to unseen but normal variations.
Multi-scale Anomaly Map Estimation:
Anomaly detection is formulated as measuring the discrepancy between the encoder features and the two reconstructed features across multiple scales. The anomaly score at each spatial location is computed using the cosine distance between the concatenated encoder features and the concatenated reconstructed features. These scale-wise anomaly maps are resized to a unified resolution, weighted, and aggregated to produce a final anomaly map, which is further refined via Gaussian smoothing.

Key Contributions

Hierarchical Encoding Strategy: The use of a frozen Swin Transformer V2 encoder to extract multi-scale representations that balance local texture details with global semantic context, specifically tailored for the MUAD setting.
Feature Diversity-Preserving Framework: The introduction of a dual-hypothesis reconstruction mechanism that mitigates deterministic feature collapse. By maintaining multiple reconstruction hypotheses, the method better preserves the diversity of normal patterns, enhancing robustness to distribution shifts and improving discrimination between normal and anomalous local structures.
Comprehensive Evaluation: The method is evaluated on three widely used MUAD benchmarks (MVTec AD, VisA, and Real-IAD), demonstrating competitive image-level performance and superior pixel-level localization accuracy compared to state-of-the-art methods.

Experimental Results
Experiments conducted on MVTec AD, VisA, and Real-IAD demonstrate the effectiveness of SwinAD:

MVTec AD: SwinAD achieves an image-level AUROC of 99.4% and a pixel-level AP of 74.4%, outperforming the strongest baseline (Dinomaly) by 5.1% in pixel-level AP, despite operating at a lower input resolution (256×256 vs. 392×392 for Dinomaly).
VisA: The method achieves a pixel-level AP of 58.0% and an F1-score of 57.3%, improving upon Dinomaly by 4.8% and 1.6% respectively.
Real-IAD: SwinAD attains the strongest image-level performance (89.5% AUROC) and the best pixel-level AP (49.4%) and F1 (52.0%) among compared methods.
Efficiency: SwinAD requires significantly fewer parameters (159.01M) and FLOPs (54.32 G) compared to Dinomaly at high resolutions, achieving approximately 20 FPS on a single NVIDIA RTX A5000 GPU.
Ablation Studies: Results confirm that intermediate encoder stages (Stage 2 and 3) provide the best trade-off for localization and that the dual-decoder diversity strategy consistently improves performance over single-decoder baselines, particularly on complex datasets like VisA and Real-IAD.

Significance and Claims
The paper claims that SwinAD provides an effective and efficient solution for multi-class unsupervised anomaly detection by combining hierarchical Swin features with diverse multi-scale reconstruction. The authors argue that the primary limitation of existing reconstruction-based methods is not insufficient reconstruction capability, but rather excessive representation collapse caused by deterministic learning. By preserving feature diversity through complementary reconstruction hypotheses, SwinAD achieves a better balance between semantic robustness and localization precision. The results indicate that this approach is particularly suitable for scalable industrial systems where computational efficiency and precise pixel-level localization are critical, offering a practical alternative to maintaining separate models for each product category.

Technical Summary: SwinAD for Multi-class Unsupervised Industrial Anomaly Detection

More like this