Multiscale Markov random fields for large image datasets representation

Abstract

Future users of satellite images will be faced with a huge amount of data. The development of content-based image retrieval algorithms for remote sensing image archives will allow them to efficiently use the upcoming databases of large images. The authors present an image segmentation and feature extraction algorithm, that will enable users to search images by content. In their approach, images are modelled by multiscale Markov random fields (MSRF). This model is superior to spatial Markov random field models in that it is able to describe the long range as well as the short range behaviour of the image data. Image information extracted at multiple scales is incorporated naturally in the model. Additionally it is computationally less costly than the spatial random field models. The difference to similar work by Ch.A. Bouman et al. (1994) and A. Jain et al. (1992) is that the multiscale process is not only used to find a reasonable segmentation of the image, but that the estimated parameters of the scale process serve also as image features. These image features together with the textural characteristics of the image are stored hierarchically in a pyramidal structure from large to small scales. Thereby even large datasets can be quickly browsed.