The idea of learning only from normal samples is straightforward to state and hard todeliver in a real factory. Two questions decide whether it works on an actual line: howmany different products one model can cover at once, and how few normal images it needs before it can start working at all. Our recent work takes up each of these in turn.
A production line rarely makes just one thing. Training and maintaining a separate model for every product does not scale, so a single model should ideally cover every category on the line at once. That is much harder than it sounds: one network must hold many different "normal" appearances simultaneously, and the reconstruction-based methods that dominate this setting carry two hidden weaknesses that get worse as categories are added. Our DDR framework identifies and repairs both.
→ Read more: Multi-Class Anomaly Detection
When a new product enters the line, waiting to collect a large set of normal images is a luxury nobody has. Few-shot detection works from a handful of images, sometimes only one, which puts enormous weight on which images those are and on where in the image the model looks for evidence. Our ReNoS framework makes both of those choices deliberate rather than accidental.
→ Read more: Few-Shot Anomaly Detection