Obtaining an estimate of the illuminant color is an important component in many image analysis applications. Due to the complexity of the problem many estrictive assumptions are commonly applied, making the existing illuminant estimation methodologies not widely applicable on natural images. We propose a methodology which analyzes a large number of regions in an image. An illuminant estimate is obtained independently from each region and a global illumination color is computed by consensus. Each region itself is mainly composed by pixels which simultaneously exhibit both diffuse and specular reflection. This allows for a larger inclusion of pixels than purely specularity-based methods, while avoiding, at the same time, some of the restrictive assumptions of purely diffuse-based approaches. As such, our technique is particularly well-suited for analyzing real-world images. On natural images, the algorithm is very stable and provides qualitatively correct estimates.