SVM KERNEL SELECTION FOR CATEGORIZATION OF IMAGES DESCRIBED BY THE BAG OF VISUAL WORDS.

This paper studies the problem of SVM kernel selection for image categorization domain. In particular, linear, RBF, chi2, histogram and Cauchy kernels were compared in terms of classification accuracy and performance. The methodology utilizes the Bag of Visual Words approach with the SIFT feature extractors to obtain the key points collection for each considered image, the k-means algorithm for visual dictionary construction and SVM for the classification process. In an experimental session subcategorization of the images depicting objects which belong to the same high level visual category is examined. In particular, the impact of different kernels on accuracy and performance of categorization is evaluated [ABSTRACT FROM AUTHOR]