Why we replaced our backbone neural network — twice
Lessons from running computer vision at the edge in dusty, vibration-heavy environments.
We have shipped three different backbone networks in Profna Vision in the last 18 months. Each replacement was painful, each was justified, and each taught us something we wished we had known earlier.
Generation one was a generic ResNet variant that worked beautifully in our lab and slowly degraded on customer floors as cameras accumulated lint and lighting drifted. Generation two was a custom CNN tuned for low-contrast textures that cut false positives by 38 percent but ran too hot inside the small steel enclosures we ship.
Generation three is a hybrid: a small transformer head sitting on top of a tuned convolutional trunk, trained with hard-negative mining on the failure cases of the previous two. It is boring, it is fast, and it survives the dust.
Talk to us
Curious how this would land in your mill?
We are happy to walk through your machines, your data, and where Profna would fit.
Book a conversation
