Faculty of Science and Technology

A seminar by DR. Edwin Lughofer from Johanes Kepler Univeristy, Linz, Austria visiting Fellow of the Royal Society (host: Dr. P. Angelov)

An On-Line Interactive and Self-Adaptive Image

Classification Framework

Edwin Lughofer

edwin.lughofer@jku.at, http://www.flll.jku.at/people/staff/lughofer.html

Johannes Kepler University Linz

Altenbergerstrasse 69, A-4040 Linz, Austria

Abstract

In many machine vision applications, such as inspection tasks for quality control, an automatic system tries to reproduce

human cognitive abilities. The most efficient and flexible way to achieve this, is to learn the task from a human expert. This

training process involves object recognition methods, adaptive feature extraction algorithms and evolving classifiers. A lot of

research has been done on each of these topics, however, simply plugging all of these methods together does not necessarily lead

to a working machine vision system.

In this talk, a generic self-adaptive image classification framework is presented, focussing on integration issues and on topics

that are specific to quality control applications. The framework is shown in Figure 1. The basic components of these framework

are the following:

• Generating contrast images by calculating the deviation images to the master

• Recognizing ROIs (=regions of interest) in the deviation images

• An adaptive object and aggregated feature extraction component (object features characterize single objects, whereas aggregated

features characterize whole images)

• Training of (initial) base classifiers for aggregated and object features (! aggregated and object classifiers) based on off-line

pre-labeled image sets

• Ensemble Classifiers for resolving contradictory input among different operators

• Strategy for on-line classification of new images (incorporating object, aggregated and ensemble classifiers)

• On-line adaptation/evolution of base and ensemble classifiers (based on operator's feedback during on-line mode)

• Early prediction of success or failure of a classifier

Each of these components will be addressed during the talk.

At the end of the talk some results on real-recorded images in various industrial systems will be presented. These results will

include a comparison of object recognition methods as well as off-line and on-line classification accuracies and achievable bounds

(from base and ensemble classifiers) based on (off-line) pre-labeled data and operator's feedback during on-line operation mode.