Color texture histograms for natural images interpretation

This paper presents a recognition method for natural images based on color texture histograms in the context of image interpretation and scene modeling. A color histogram of sums and differences is proposed to obtain texture features which are faster to compute than correlograms ( i.e., colored version of co-occurrence matrices) and improving substantially object recognition. Outdoor natural images are generally affected by color casting artifacts which can affect object recognition. Therefore, an on-line color balancing algorithm based on chromatic adaptation models, eliminates these color deviations. The proposed approach globally involves functions as color segmentation, histogram texture analysis and a region recognition step. Our approach has been extensively tested and validated to obtain an accurate 2D scene interpretation from natural images. This technique may be used in robot navigation by identifying navigable regions ( e.g., roads or fairly flat surfaces) on natural scenes, scene modeling and image categorization.