Dark channel based illumination invariant feature detection

This paper provides a novel feature detection method which utilizes illumination invariant space to achieve a high performance of robustness under the variating lighting conditions. Taking advantage of the dark channel prior knowledge, the proposed method builds three indicators to describe the illumination invariant components in the RGB color space and eliminates the light sensitive parts. The components retained are transformed to the illumination invariant space in which the traditional feature detection methods works more robustly. In contrast to the current transformation method, the method gives a clearer projection from the RGB space to the illumination invariant space which improve the discerning ability of the feature detection methods. The dark channel prior knowledge helps not only the building of more distinguishable indicators, but also the detection of edge features of an object.