Title: A Flexible Framework for Learning Visual Quality Control from Human Operators
Authors: Sannen, Davy
Van Brussel, Hendrik
Issue Date: May-2009
Host Document: Proceedings of 4th International Conference on Optical Measurement Techniques for Structures and Systems, Abstract Book
Conference: International Conference on Optical Measurement Techniques for Structures and Systems edition:4 location:Antwerp date:25-26 May 2009
Abstract: The most flexible and effective way to reproduce the human cognitive abilities needed to automate the complex decision tasks in production processes, such as visual quality control, is by learning these tasks from human experts. This paper presents a novel image classification framework in which the human operators, who currently perform the task by hand, can directly train the system. Special emphasis is placed on the general applicability of the framework to arbitrary surface inspection problems.

The different processing steps within this framework are highlighted, starting from a generic interface to remove application-specific elements of the images produced by the vision system towards the upstream processing steps, in the form of so-called "deviation images", indicating potential defects in the images. After the identification of relevant Regions Of Interest (ROIs), the images are described using different generic features which can be organised in a hierarchical form, depending on the users' input. For instance, an operator can classify an entire image as being defective or could provide more detailed information about each of the identified ROIs. It is shown how these different levels of input can be integrated into the learning system. A number of trainable classifiers (including decision trees, Bayesian and clustering-based classifiers) is evaluated and it is shown how a more detailed user input can improve the predictive accuracy of these classifiers. Finally, ways the system can cope with input from different, possibly contradicting operators are highlighted. This is done using so-called ensemble techniques, which integrate the predictions of a set of classifiers into a combined decision.

Classification results are presented for a number of different visual surface inspection tasks, including inspection of CD imprints, eggs, rotors and bearings. The improvement of the predictive accuracy is shown which can be achieved by incorporating the different levels of user input for these applications.
Publication status: published
KU Leuven publication type: IMa
Appears in Collections:Production Engineering, Machine Design and Automation (PMA) Section

Files in This Item:
File Description Status SizeFormat
OPTIMESS2009Abstract.docAbstract Published 39KbMicrosoft WordView/Open


All items in Lirias are protected by copyright, with all rights reserved.