An adaptive method for natural scene analysis

Abstract

In the field of artificial vision, researchers have almost always focused their attention on analysis and processing of monochromatic images, restricting the use of colours to the display phase, which involves an interaction between the human user and the machine. In recent years, acquisition and processing systems for colour images have became increasingly available or accessible, so that this research area is now under continuing development and rapidly growing in importance.

This paper deals with analysis of natural scenes, represented by colour images. In particular, low-level processing steps are investigated, from filtering to segmentation, while, for the moment, reduction in information with feature-selection techniques is not considered. The proposed filtering utilizes a widely used (in the monochromatic case) edge-preserving smoothing filter which has been extended to the multichromatic case.

Concerning the segmentation process, a classical region-growing approach is first followed, and, subsequently, an adaptive multistep processing algorithm is adopted, which is based on both fuzzy techniques and a backtracking procedure to obtain a set of regions that compare favourably with the original image.