Semantic image segmentation with a multi-dimensional Hidden Markov Model

Segmenting an image into semantically meaningful parts is a fundamental and challenging task in image analysis and scene understanding problems. These systems are of key importance for the new content based applications like object-based image and video compression. Semantic segmentation can be said to emulate the cognitive task performed by the human visual system (HVS) to decide what one "sees", and relies on a priori assumptions. In this paper, we investigate how this prior information can be modeled by learning the local and global context in images by using a multidimensional hidden Markov model. We describe the theory of the model and present experiments conducted on a set of annotated news videos.