SUBMITTED TO THE IEEE TRANSACTIONS ON NEURAL NETWORKS , 1999 1 SVMs for Histogram - Based ImageClassi

Traditional classiication approaches generalize poorly on image classiication tasks, because of the high dimensionality of the feature space. This paper shows that Support Vector Machines (SVM) can generalize well on diicult image classiication problems where the only features are high dimensional histograms. Heavy-tailed RBF kernels of the form K(x; y) = e ? P i jx a i ?y a i j b with a 1 and b 2 are evaluated on the classiication of images extracted from the Corel Stock Photo Collection and shown to far outperform traditional polynomial or Gaussian RBF kernels. Moreover, we observed that a simple remapping of the input x i ! x a i improves the performance of linear SVMs to such an extend that it makes them, for this problem, a valid alternative to RBF kernels. IV Average error rates on Corel14. Each columns corresponds to a diierent kernel. The rst line reports the average number of support vectors required for the full recognizer (i.e. 14 \one against the others" SVM classiiers). The next lines report the error rates using nonlinear input remappings (exponentiation by a). VII Class-confusion matrix for a = 0:25 and b = 1:0. For example, row (1) indicates that on the 386 images of the Airplanes category, 341 have been correctly classiied, 22 have been classiied