Uploaded November 2026 | Updated November 2026, 1 hour ago
See how computer vision anomaly detection can catch defects like burnt cookies in a food processing line, even when you don’t have examples of the burnt cookies included in your training data.
A major challenge when using computer vision to catch defects, damage or other problems in manufacturing is finding the example images to train an object detection model. Anomaly detection flips the situation so you can train a computer vision model on images of what is considered normal, and then the model will flag images that appear to be outside the norm.
In this video, Gustavo Reis from Roboflow shows an example project where he trained an anomaly detection model to look at a cookie line and find the cookies that are burnt. He walks through the process of choosing a model type, PatchCore or FoundAD, building a Workflow to visualize results and ignore false positives, and testing it on his footage.
= Chapters =
0:00 Intro: Catching Burnt Cookies With Anomaly Detection
0:47 Why Train on Normal Footage Instead of Labeling Defects
1:58 Create an Anomaly Detection Project in Roboflow
2:55 PatchCore vs FoundAD: Which Anomaly Model to Train
3:40 Build the Anomaly Detection Workflow: Heatmap and Threshold
4:18 Results: Very Burnt, Burnt and Barely Burnt Cookies
= Resources =
Anomaly detection launch post, with PatchCore vs FoundAD benchmarks: roboflow.com/blog/roboflow-anomaly-detection
Anomaly detection docs: docs.roboflow.com/models/supported-tasks/anomaly-detection
Roboflow Agent: docs.roboflow.com/agents/roboflow-agent
Vision Events: docs.roboflow.com/deployment/monitoring-and-analytics/vision-events
See how computer vision anomaly detection can catch defects like burnt cookies in a food processing line, even when you don’t have examples of the burnt cookies included in your training data.
A major challenge when using computer vision to catch defects, damage or other problems in manufacturing is finding the example images to train an object detection model. Anomaly detection flips the situation so you can train a computer vision model on images of what is considered normal, and then the model will flag images that appear to be outside the norm.
In this video, Gustavo Reis from Roboflow shows an example project where he trained an anomaly detection model to look at a cookie line and find the cookies that are burnt. He walks through the process of choosing a model type, PatchCore or FoundAD, building a Workflow to visualize results and ignore false positives, and testing it on his footage.
= Chapters =
0:00 Intro: Catching Burnt Cookies With Anomaly Detection
0:47 Why Train on Normal Footage Instead of Labeling Defects
1:58 Create an Anomaly Detection Project in Roboflow
2:55 PatchCore vs FoundAD: Which Anomaly Model to Train
3:40 Build the Anomaly Detection Workflow: Heatmap and Threshold
4:18 Results: Very Burnt, Burnt and Barely Burnt Cookies
= Resources =
Anomaly detection launch post, with PatchCore vs FoundAD benchmarks: roboflow.com/blog/roboflow-anomaly-detection
Anomaly detection docs: docs.roboflow.com/models/supported-tasks/anomaly-detection
Roboflow Agent: docs.roboflow.com/agents/roboflow-agent
Vision Events: docs.roboflow.com/deployment/monitoring-and-analytics/vision-events










