Conditional Random Fields for High-Level Part Correlation Analysis in Images Giuseppe Passino and Ebroul Izquierdo Queen Mary, University of London Mile End Rd London, E1 4NS, UK {giuseppe.passino,ebroul.izquierdo}@elec.qmul.ac.uk Abstract. A novel approach to model the semantic knowledge asso- ciated to objects detected in images is presented. The model is aimed at the classification of such objects according to contextual information combined to the extracted features. The system is based on Conditional Random Fields, a probabilistic graphical model used to model the condi- tional a-posteriori probability of the object classes, thus avoiding prob- lems related to source modelling and features independence constraints. The novelty of the approach is in the addressing of the high-level, se- mantically rich objects interrelationships among image parts. This paper presents the application of the model to this new problem class and a first implementation of the system. 1 Introduction The part-based image analysis is a promising approach for object detection and image classification systems, being based on the strong correlation among objects represented within images. A natural way to address this problem is represented by the probabilistic graphical models, in which a node in a graph is associated to each image part in order to perform inference on them (i.e., deduce a la- belling). Such models make the relationships between nodes explicit, simplifying the inference and giving to the problem an immediate interpretation. The Con- ditional Random Fields (CRF) [1] represent a solution to model the part labels probabilities and the dependencies between the extracted features and the part labels configuration. The system presented in this paper is based on the idea that a CRF can be applied to study high-level relationships between semanti- cally rich atomic objects. As the characteristic dimension of the parts considered as constituents of the image to be labelled grows, two effects are expected: the reduction of the number of elements to handle, and possibly the growth of the semantics associated to the elements. This application is conceptually rather different from the original problems solved using CRF models, and an analysis and reconsideration of major assumptions has to be made. B. Falcidieno et al. (Eds.): SAMT 2007, LNCS 4816, pp. 264–267, 2007. c  Springer-Verlag Berlin Heidelberg 2007