How to Build AI Visual Inspection System for Visual Defect Detection in Manufacturing

[@WurstPeterl]

The real problem is that the crack on the bottle is moving.

[@oleksiitsymbal8205]

the analysis could be done from a photo, extracted out of a stream. But yes - as its moving, it should be taken into account while notifying

[@WurstPeterl]

Oleksii Tsymbal I meant that the crack is obviously added in post to stock footage. During the zoom out you can see the crack is moving on the bottle.

[@WurstPeterl]

It was meant to be a joke and is obviously not a real concern

[@oleksiitsymbal8205]

@@WurstPeterl now got it ))

[@ItsNotAllRainbows_and_Unicorns]

Good catch, Yep cracks in the plastic (which don't appear like that on plastic and more on cement) would have to be caught before placing on the bottle. CV would be best used on bottle labels, fluid levels and improper placement of the caps or missing caps. It's really up to the vendor that sells the plastic caps to detect anomalies and not the bottler. BTW, that fluid looks like piss and the presenter in the video looks very uncomfortable talking to the camera. Odd video.

[@eugy2008]

Can this system detect the exact same crack in different settings? For example, say the bottle moves from one warehouse (room) to another, would the system be able to tell the same crack it saw already in another location or would it just identify the crack it saw before just another crack?

[@MobidevBizCorp]

If I got it right, yes. Once the model receives enough data for training, it has no problem in detecting the crack and understanding that this is the same crack it saw few minutes ago.

[@snadeepkokare]

is there way to detect some unwanted things from bunch of materials . like some unknown object which are not listed in our dataset .

[@MobidevBizCorp]

Sure, the model could be trained to detect different objects, and then notify only if some of the objects are unknown to it. Anyway, each case should be evaluated in order to give you a more precise answer. If you have photo samples, it will help engineers to say if it is possible to detect or not.

[@ayanshaikh3550]

What software and hardware is required in this process

[@MobidevBizCorp]

It depends on the scale of a system and exact tasks you want to complete. Usually, you'll need computer vision cameras as a video recording device, and a server where the model will operate. That's what you need concerning hardware as a bare minimum. In terms of software, well. Neural network is required to fetch the features and apply them to launch detection and classification tasks. Then you'll need a computer vision model to detect anomalies on the image. Besides this there is a machine learning pipeline that serves training tasks and all burden corresponding data transmission, transformation, cleaning, etc.