Experiments with an Adaptive Hidden Markov Mesh Image Model

This paper is concerned with the segmentation of images from an image sequence representing a dynamic scene. We assume a rapid scene sampling producing a situation similar to the short range motion process of the human visual system. Image segmentation is the division of the image into different regions, each having certain properties. Here, as in a number of our previous works with static imagery, we adopt a Markov Mesh random field model for the contextual information hidden in-the (unobservable) distribution of the region labels over the image pixels. A clustering method is used for estimating the model parameters and grouping the pixels in feature space into clusters. The clusters are then mapped back to the original spatial domain on a maximum a posteriori probability basis thereby producing a segmentation of the image. Our way of processing images sequences capitalizes on the facts that under the short range motion hypothesis, a frame transition produces relatively minor perturbations of the gray-level histogram and model parameters, and the outcome of our segmentation algorithm is not very sensitive to minor perturbations of the optimal model parameters. Accordingly, our approach combines the segmentation method developed previously with a frame by frame updating of the parameters for the hidden Markov mesh image model. Experimental results with an image sequence from a real dynamic scene provide evidence that our approach is subjectively relevant to the dynamic segmentation problem.