Learning to Label Images
7The problem of image labeling, in which each pixel is assigned to one of a finite set of labels, is a difficult problem, as it entails deciding which components of an image belong to the same object as well as classifying the components. I will describe two approaches we have taken to this problem, both utilizing conditional random fields to model contextual effects. The first uses a novel form of learning higher-order structure, which we developed for this work but has broader applicability. The second is a simpler and more efficient method that turns out to work just as well. In both cases, the model is trained on a database of images and the learning method estimates model parameters by maximizing a lower bound of the data likelihood. We examine performance on three real-world image databases, and compare our system to a standard classifier and other conditional random field approaches.