Real - Time Real - World Visual Classi cation { Making Computational Intelligence

An Intelligent Inspection Engine (IIE) for classiication of non-regular shaped objects from images is described and evaluated using real-world data from a waste package sorting application. The entire system is self-organizing. Principal component analysis and additional a priori knowledge on color properties are used for feature extraction. As classiiers growing neural networks provide robustness and minimize the number of runs for parameter tuning. We propose a method to encompass feature extraction and classiication within a bootstrap procedure. These method reduces the immense memory requirement for the computation of principal components if number and size of training images are huge without to much loss of recognition quality.