Machine Learning in Machine Vision

The word “Machine learning” is used when a machine imitates "cognitive" functions, often associated with humans such as “learning” or “problem solving”.

Read about the very basics of machine learning.

Machine learning is often implemented via a (convolutional) neural network and can be supervised or unsupervised.

Supervised learning can provide solid results and is often considered the best approach, but it requires many annotated training images. Annotation is time-consuming and it can be difficult to get images upfront in a machine vision application. 

JLI vision works with "Hybrid Vision" which we consider to be a combination of: 

• Machine learning 
• Traditional machine vision 
• Craftsmanship.

Structural applications such as measuring dimensions and depth can be performed by traditional machine vision. They are easy to test and potentially expand with functionality.

Hybrid vision is the perfect solution for "aesthetic" applications. Aesthetic applications consist of ensuring that the finished product fulfills the end-users or consumers' expectations. Typical defects could be scratches, holes, or color deviations.

As opposed to structural applications it can be difficult to define objective requirement specifications.

In practice good and bad samples are collected and used as references in manual inspection.

Previously aesthetic applications were often left unsolved because of the shortcomings of traditional machine vision, but with hybrid vision it is in some cases possible.

Application example: Wood inspection

• Objective: Detection of gnarls and resin pockets in glue boards
• Solution: Glue Boards are scanned on a conveyor and processed using a combination of machine learning (determines if the candidate is defective), traditional machine vision (selects candidates), and 3D (checks the surface of glue boards). 
• Benefit: Solves a time-consuming manual inspection task inline in production and enables fully automatic repair of glue boards.
• Result: Achievable accuracy +95% which is far better and more consistent than manual inspection.

Application example: Glass inspection

• Objective: Detection of open- and closed end airlines in glass tubing
• Solution: Using traditional machine vision all airline defects are detected. Defects are then in real-time processed by machine learning in order to determine whether the defect is open- or closed end. 
• Benefit: Improves yield by reducing scrap.
• Result: Achievable accuracy +95% 
  
  

Application example: Steel inspection

• Objective: Detection of freak defect on the surface of rails directly after production
• Solution: Rails are scanned by passing through an inspection tunnel. Images are processed real time using a combination of machine learning, traditional. Machine learning network is taught unsupervised.
• Benefit: Solves a time consuming manual inspection task inline in production.
• Result: Achievable accuracy +90% which is better and more consistent than manual inspection

Video: Automating complex inspection tasks

Learn more about how we work with machine learning in this episode of our video series, The Vision Lab: "How can you automate complex aesthetic inspection tasks?"