CORDIS - EU research results

Dynamically Reconfigurable Quality Control for Manufacturing and Production Processes Using Learning Machine Vision

Final Report Summary - DYNAVIS (Dynamically Reconfigurable Quality Control for Manufacturing and Production Processes Using Learning Machine Vision)

Inspection of products by machine vision often has to solve the problem of how to implement a human decision-making process in software. Currently, this requires a step-by- step reprogramming or parameterisation of the software, which may take very long. The results of this project will enable us to use human-machine cooperation to learn complicated inspection tasks instead of step-by-step improvements and adaptations of software. The system will automatically adapt to specific (or changing) requirements.

The 'Dynamically reconfigurable quality control for manufacturing and production processes using learning machine vision' (Dynavis) project focused on the development of 'trainable' machine vision algorithms and of appropriate machine learning techniques. In order to create such methods they focused on the following scientific objectives:

- machine learning methods for processing the complicated data produced by the vision system
- methods to deal with multiple, possibly contradictory input by the operators
- methods for predicting success or failure of the learning process in early stages of the training process.

Dynavis will enable the machine vision system to directly learn from the human operator. Based on this input the machine vision system will gradually build a hypothesis about which parts are good and which are bad.

Two hardware demonstrators were set up and also shown at the final meeting. The first one was the rotor scanner. This demonstrator was built early in the project and used to scan images of rotors at Atlas Copco. In the last year of the project it was converted to a demonstrator that also processes the images, calculates features and classifies the images with a set of trained classifiers and ensembles.

The second demonstrator was based on a CD print inspection system that checks the printed side of the CD for any kind of printing defects. A hardware setup consisting of a light source and camera was built and integrated with the Dynavis software. The classifiers were trained on existing data sets and could be shown to improve the decisions made by the original inspection software, in particular by recognising so-called pseudodefects.