Architecture for image labelling in real conditions

Abstract

A general model for the segmentation and labelling of acquired images in real conditions is proposed. These images could be obtained in adverse environmental conditions, such as faulty illumination, non-homogeneous scale, etc. The system is based on surface identification of the objects in the scene using a database. This database stores features from series of each surface perceived with successive optical parameter values: the collection of each surface perceived at successive distances, and at successive illumination intensities, etc. We propose the use of non-specific descriptors, such as brightness histograms, which could be systematically used in a wide range of real situations and the simplification of database queries by obtaining context information. Self-organizing maps have been used as a basis for the architecture, in several phases of the process. Finally, we show an application of the architecture for labelling scenes obtained in different illumination conditions and an example of a deficiently illuminated outdoor scene. © Springer-Verlag Berlin Heidelberg 2003.