Color-Contrast Landmark Detection and Encoding in Outdoor Images

Abstract

This paper describes a system to extract salient regions from an outdoor image and match them against a database of previously acquired landmarks. Region saliency is based mainly on color contrast, although intensity and texture orientation are also taken into account. Remarkably, color constancy is embedded in the saliency detection process through a novel color-ratio algorithm that makes the system robust to illumination changes, so common in outdoor environments. A region is characterized by a combination of its saliency and its color distribution in chromaticity space. The newly acquired landmarks are compared with those already stored in a database, through a quadratic distance metric of their characterizations. Experimentation with a database containing 68 natural landmarks acquired with the system yielded good recognition results, in terms of both recall and rank indices. However, the discrimination between landmarks should be improved to avoid false positives, as suggested by the low precision index.