Scene Categorization Based on Extended Color Descriptors

In automatic annotation systems, a scene categorization involves the compulsory stage of descriptor extraction in order to build a histogram of visual words. A family of new color descriptors based on point features, which are invariant not only to geometric transforms but also light changing, is investigated. In following, the algorithm executes a preliminary color and texture segmentation based on J-SEG algorithm. The received regions are ranked by areas. The extended color descriptors computing in 5–7 large area regions are applied for visual word construction. Then images are categorized by support vector machine. The comparative results of experimental estimators present the precision values of image categorization by use a test dataset containing 2,688 images.