machine vision 2-min.jpg

Visual inspections with AI - our use cases

Deevio manages to bring visual inspections to the next level by applying artificial intelligence. We can show from already successfully implemented projects that we are industry agnostic and can automate entire systems as well as retrofit existing systems.

Automotive component manufacturers are subject to strict production requirements and guidelines that leave no room for error when it comes to defect detection. 

The natural variance of the raw materials and the difficulties during the die casting process result in scrap rates between 5-10% with defects such as scratches, blowholes, or excess material.

Use Case

Our customer is a German aluminum die casting company that produces more than 600.000 engine parts for Tier 1 automotive suppliers every year (until now visual quality control is made manually).

  • Defined the appropriate image acquisition setup to capture all the different defects in the images

  • Acquired images of OK and NOK parts and labelled defects 

  • Developed an AI segmentation model that can detect all defects and is able to cope with the surface variations of the parts

automotice-min.jpg
 
precision tooling2-min.jpg

Precision tools are an integral part of CNC machines that are used to make critical parts in many industries such as oil and gas, space, and automotive.

Previous attempts to automate this visual inspection task failed due to the small size and the different shapes of defects, such as scratches, broken edges, or dents. Moreover, the high product variety of precision tools makes it hard to effectively deploy traditional rule-based machine vision software which would need to be updated and configured for each type of product.  

What we did:

  • Defined the appropriate image acquisition setup to capture the different defects in the images on a micron-level

  • Defined the industrial an automation setup needed to handle the parts and feed them to the image acquisition system

  • Developed an AI segmentation model that can detect all defects and works with all different shapes of the precision tools

 

Similar to vial production, quality requirements for the pharmaceutical packaging in the pharmaceutical industry are extraordinarily high. Manufacturing companies use machine vision systems to detect defects such as scratches, holes, or deformations. The machine vision systems that are currently used in the industry are good at detecting defects, but at the same time, they have high false rejection rates that result in high costs and waste for manufacturing companies.

Use case:

  • Retrofitted an existing machine vision system by connecting our AI box to the image acquisition setup for data transfer of up to 30 images per second 

  • Labelled OK and NOK images together with the customer’s domain experts

  • Developed an AI segmentation model that reliably detects NOK parts as NOK parts while reducing the false eject rate to < 1%

packaging, plastic bottles-min.jpg
 
Vaccine vials.jpg

In the pharmaceutical industry, automated visual inspection (AVI) systems are already industry standard. AVI systems are more consistent, more reliable, and faster than human inspectors. However, in some applications such as vial inspection AVI systems produce up to 20% of false positives (defined as parts that are OK, but misclassified by the AVI system as NOK). This results in annual costs for the industry of up to EUR 750mn. 

 

One typical cause for this misclassification is the presence of water bubbles in the vial (which are ok) that the AVI recognizes as black dots (which are NOK).


Deevio’s AI software can reliably separate water bubbles from black dots and solve this problem while maintaining the false reject rate below  1%.

What we did:

  • Retrofitted an existing AVI system by connecting our AI box to the image acquisition setup for data transfer of up to 70 images per second

  • Labeled OK and NOK images together with the customer's domain experts

  • Developed an AI model that reliably detects NOK parts as NOK parts while reducing the false eject rate to < 1%