Description:
At the seminar, the concept of texture implicitly will be investigated. It suggests some local shift invariant statistical properties. Motivated by this idea we have shown in a series of papers that grey-scale textures can be modeled by estimating the joint probability density of grey levels in a suitably chosen observation window. The texture parts chosen by shifting the window (in a vector arrangement) can be viewed as observations of a random vector identically distributed with an unknown joint probability density. The method is based on estimation of the unknown probability density in the form of a normal mixture of product components by means of EM algorithm.
The method can be applied e.g. to identify defects or abnormalities in grey-scale texture images and also it can be used to evaluate screening mammograms. In view of the fact that most mammograms are pathology-free the malignant abnormalities should be identifiable via novelty (outlier) detection. In this way many diagnostically important details become visible and easy to identify.