Spectral gradients for color-based object recognition and indexing

Abstract

In the past decade color has been developed as an effective cue for object recognition and image indexing. The initial draw of color was how easily object signatures could be captured and compared. Simple histograms of colors radiated from an object proved to be a stable representation of many objects without requiring complex geometric processing to address object pose. While recognition/indexing should depend only on colors intrinsic to object reflectance, a color signature computed from raw image colors is highly sensitive to the color and pose (direction) of incident illumination. Several methods have explored illumination-color invariants, and most of the recent invariants to illumination color and pose employ some form of chromacity representations based on the ratios of image color values or their combinations. This paper presents a comprehensive framework for logarithmic chromacities for the interpretation of image color change due to illumination pose and color. A log-chromacity space is developed from spectral (color) derivatives of logarithmic image irradiance, called spectral gradients, to achieve invariance to global illumination color and pose. In this space, the multiplicative influence of illumination color and pose on image irradiance become additive effects, and this facilitates the analysis of illumination color and pose. A representation from spatial derivatives of spectral gradients, called spectral–spatial gradients is also developed to additionally provide invariance to local variation of illumination color within an image. Experimental results are presented to demonstrate the efficacy of the proposed descriptors.