Image categorization based on segmentation and region clustering

The image categorization problem is about assigning images to a set of predefined categories or classes. Machine learning algorithms can be used to learn models that will assign images to classes, but most machine learning algorithms cannot work with images directly. Therefore, we need to introduce a representation of images that typical learning algorithms can use. We propose two image representation methods based on texture segmentation. In the region clustering approach, segmentation is applied to all images in the training set. The descriptions of the resulting regions are then clustered, and images are represented by sparse vectors where each component indicates whether, and to what extent, regions from a particular cluster are present in the image. Algorithms such as Support Vector Machines can then be used to train on the resulting vectors. Alternatively, a similarity measure between segmented images may be taken as a starting point and converted into a generalized kernel for use with Support Vector Machines. We compare these two approaches to a more basic representation based on autocorrelograms.