AI xPress

AI and computer science research.

Back to Foragentis

Machine Learning Methods

EfficientAD flags visual anomalies in two milliseconds

Machine Learning Methods October 31, 2026

Light trails branching at a highway interchange at night

In a 2024 paper at the IEEE Winter Conference on Applications of Computer Vision, researchers at the Technical University of Munich describe EfficientAD, a fast detector for visual anomalies. Anomaly detection means finding images that differ from normal examples. EfficientAD flags unusual patterns in about two milliseconds and processes six hundred images per second. Across 32 datasets from three industrial anomaly detection collections, the authors report improved results for both detecting anomalies and marking where they are in an image.

How does EfficientAD find anomalies?

EfficientAD uses a student-teacher setup. A student network learns to predict features, small visual patterns a network extracts, from normal training images that contain no anomalies. At test time, the student fails to predict features of anomalous images. That failure marks the image as unusual. A training loss, a penalty used during learning, keeps the student from imitating the teacher too well beyond the normal examples. The authors say this cuts computing cost while improving detection.

How does EfficientAD catch logical anomalies?

Logical anomalies are invalid combinations of normal local parts, such as objects in the wrong order. EfficientAD adds an autoencoder that analyzes the entire image to catch these global mistakes. The authors say this route handles the challenge efficiently.

Frequently asked questions

How fast is EfficientAD?

The paper reports a latency of two milliseconds and a throughput of six hundred images per second. The lightweight feature extractor alone processes an image in less than a millisecond on a modern GPU.

What data did the researchers use?

They tested EfficientAD on 32 datasets from three industrial anomaly detection dataset collections.

What do the authors claim about the results?

The authors report that EfficientAD set new standards for both anomaly detection and localization. They describe the combination of low error rate, low latency and high throughput as an economical solution for real-world use and a basis for future research.

Limits and what's next

The abstract describes one study. The researchers tested EfficientAD on 32 datasets from three industrial anomaly detection dataset collections. The setting is a benchmark evaluation on a modern GPU. The abstract reports no independent replication. It reports a low error rate but does not state the exact error figure. The authors describe the work as a basis for future research.

Source: Kilian Batzner et al., “EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies,” IEEE Winter Conference on Applications of Computer Vision, 2024.