Machine Vision is widely used for metrology purposes, but lately more and more applications require an algorithm that can detect aesthetic defects. Algorithms like that can be really difficult to break down, from image perception to algorithm.

Machine vision

Machine Learning (ML) is ideal, to extract image based features and make a decision based on all features. Convolutional Neural Networks (CNN’s) can be trained on large sample sets to achieve filters, that extract the important image features needed to classify the images or detect defects.

 

Rule of thumb:

If the defect is visible in the image to the human eye, it can be detected by ML.

 

Machine learning:

Traditionally Machine learning has required a large amount of annotated images for supervised learning. But the workload in annotating images, is not feasible.

Machine Learning techniques like, Transfer Learning and Unsupervised Learning, are much more feasible and can handle inspection demands in an agile production environment.

Transfer Learning, uses networks, designed for similar purposes, as a starting point, and can quickly achieve high accuracy from a small annotated sample set.

Unsupervised Learning, is based on unsupervised samples. Auto Encoders can learn features from the input image itself, and then fine tuned with a small set of annotated images.

 

Applied Unsupervised Learning:

Unsupervised samples can also be used to visualize clusters, based on features, that may not be obvious when looking at the image. These clusters, can be output classes for classification purposes. Below is a figure with 5 clusters, based on extracted image features from unsupervised images and visualized using K-means clustering.

A unsupervised sample set of digit images would result in 10 clusters.

target clusters

 

New call-to-action

 

Future Machine vision Machine learning Deep learning Vision system