Learning color names using least-squares probabilistic classifiers

Abstract

Color name is one of the important features for computer vision. Many existing methods proposed to classify colors into a small number of color names. In this paper, we propose an alternative method with the goal to improve the accuracy for assigning a color name to an object in the given image. We here use the least-squares probabilistic classifiers (LSPC) with the local scaling parameters for solving this task. The benefit of the LSPC is that its solution can be computed analytically so that the obtained solution is global optimum, while the local scaling parameters play an important role to deal with the data including clusters with different local statistics as appeared in the real-world data. To deal with this task, the LSPC is learned to assign a color name to each pixel with the highest of the class-posterior density distribution. Then, the estimations of the class-posterior density distributions are utilized to compute the scores for predicting a color name to the given object. Lastly, the color name with the highest score is chosen as a predicted color name for that object. The experimental results on the eBay data set show the improvements over previously proposed
methods.