A physics-motivated approach to detecting sky in photographs

Sky is among the subject matters frequently seen in photographs and useful for image understanding, processing, and retrieval. We propose a novel approach to sky detection based on color classification, region extraction, and physics-motivated sky signature validation. First, the color classification is performed to generate a belief map of sky colored pixels. Next, connected components are extracted from the sky color belief map to generate candidate sky regions. Finally and most importantly, we determine the orientation of a candidate sky region, analyze traces within the region based on a physics-motivated model, and compute the sky belief of the region. For a database of 1800 amateur photos of variable content and quality, the recall rate is 96% with a precision rate of 98% on a per region basis.