Using image segmentation as a basis for categorization

Image categorization is the problem of classifying images into one or more of several possible categories or classes, which are defined in advance. Classifiers can be trained using machine learning algorithms, but existing machine learning algorithms cannot work with images directly. This leads to the need for a suitable way of representing or describing images such that learning algorithms can work with them. We consider a representation based on texture segmentation and a similarity measure between segmented images which has been used successfully in the related area of image retrieval. A generalized kernel for use with the support vector machine (SVM) algorithm can be built from such a similarity measure. We compare this approach with a more straightforward representation based on autocorrelograms, and we show that these two representations can be combined to obtain classifiers with higher categorization accuracy.