Robust object extraction with illumination-insensitive color descriptions

The color histogram is a powerful cue for object extraction and recognition. Local color histogram matching is robust to changes in location, and size and to complex backgrounds, and can be sped up considerably by the active search method using efficient upper bound pruning. This strategy, however, has significant limitations in the event of changes in illumination. In an effort to overcome these limitations, we have focused on color ratios. This paper describes an extended histogram for focused (focal) color matching. Called a weighted histogram, it is based on color ratios from neighboring locations and evaluates each of them individually according to their significance. This histogram is relatively insensitive to illumination changes and can reduce the influence of noise or quantification errors. We present a series of experiments that show the effectiveness of this weighted histogram matching with real images taken under different illuminations.