From Salient Features to Scene Description

In this paper, we discuss an image modelling method that can capture and represent a number of commonly used image characteristics such as lines, blobs, statistical and structural textures. We develop the concepts underlying the Kadir and Brady feature detection algorithm [2], namely feature space unpredictability and spatial localisation, to characterise, what is termed, the semilocal predictability of features. Many previous approaches to image description choose a particular representation a-priori. In contrast, the proposed approach aims to detect the presence of particular types of region. We suggest that the result is a richer description of the image from which semantic scene content can be extracted. Preliminary results are presented.