Semantik renk degişmezliği

Color constancy aims to perceive the actual color of an object, disregarding the effect of the light source. Recent works showed that utilizing the semantic information in an image enhances the performance of the computational color constancy methods. Considering the recent success of the segmentation methods and the increased number of labeled images, we propose a color constancy method that combines individual illuminant estimations of detected objects which are computed using the classes of the objects and their associated colors. Then we introduce a weighting system that values the applicability of the object classes to the color constancy problem. Lastly, we introduce another metric expressing the detected object and how well it fits the learned model of its class. Finally, we evaluate our proposed method on a popular color constancy dataset, confirming that each weight addition enhances the performance of the global illuminant estimation. Experimental results show promising results, outperforming the conventional methods while competing with the state of the art methods.