Markov random fields and spatial information to improve automatic image annotation

Content-based image retrieval (CBIR) is currently limited because of the lack of representational power of the low-level image features, which fail to properly represent the actual contents of an image, and consequently poor results are achieved with the use of this sole information. Spatial relations represent a class of high-level image features which can improve image annotation. We apply spatial relations to automatic image annotation, a task which is usually a first step towards CBIR. We follow a probabilistic approach to represent different types of spatial relations to improve the automatic annotations which are obtained based on low-level features. Different configurations and subsets of the computed spatial relations were used to perform experiments on a database of landscape images. Results show a noticeable improvement of almost 9% compared to the base results obtained using the k-Nearest Neighbor classifier.