Trained texture segmentation using data mining algorithms

Texture based segmentation is a topic where a lot of different approaches lead to more or less satisfying results. In general all of them try to match a particular feature or a feature vector which describes the analyzed region. Subsequently a threshold or threshold vector is applied and a texture class is assigned to the region. This paper describes how data mining algorithms can be used advantageously for texture based segmentation. Using a reference image with known texture, a model for a classi er is trained, that is applied to image regions of unknown texture. For the data mining it is necessary to calculate many different features and rate them (e.g. by their information gain or correlation) accordingly. Only the best features selected this way are used to train a classi er, which is then used to segment subsequent images. Using this selected classi er, it is possible to determine the location where a sped c texture occurs in the image. The performance of the classi er is demonstrated for synthetic test images and the problem of detection of scratches on a metal sheet under inhomogeneous illumination. In this example only two reference images are classified manually to train the classi er and the rest is done automatically. So, no additional parameters or thresholds must be set for the scratch detection problem analyzed.