Improving visual recognition using color normalization in digital video applications

Color variations in video imagery caused by imperfections of acquisition devices from analog sources severely penalize the performance of visual recognition tasks, which are very important for smart video indexing, annotation and browsing, for instance. We propose a color correction technique based on modeling the distribution in RGB space of the colors generated by a particular device using a mixture of Gaussian probability density functions, in order to automatically obtain transformations that let us have a normalized representation of color images. It has been tested within the framework of appearance based recognition of video key-frames. In this case, the representation of key-frames consists of their projection into a low dimensional linear subspace given by principal component analysis (PCA), where the matching is done using minimum Euclidean distance. The application of color correction significantly improves the recognition rate, giving nearly optimum results