Motion compensated color image : classification and parameter : estimation in a Markovian framework

This paper deals with the classification of color video sequences using Markov Random Fields (MRF) taking into account motion information. The theoretical framework relies on Bayesian estimation associated with MRF modelization and combinatorial optimization (Simulated Annealing). In the MRF model, we use the CIE-luv color metric because it is close to human perception when computing color differences. In addition, intensity and chroma information is separated in this space. The sequence is regarded as a stack of frames and both intra- and inter-frame cliques are defined in the label field. Without motion compensation, an inter-frame clique would contain the corresponding pixel in the previous and next frame. In the motion compensated model, we add a displacement field and it is taken into account in inter-frame interactions. The displacement field is also a MRF but there are no inter-frame cliques. The Maximum A Posteriori (MAP) estimate of the label and displacement field is obtained through Simulated Annealing. Without parameter estimation, our model would not be useful in real-life applications. We propose herein a method to compute model parameters on the first frame (regarded as a still color image) of the input video sequence. The proposed method estimates mean vectors effectively even if the observed image is very noisy and the histogram does not have clearly distinguishable peaks. These values are then used in a more complex, iterative estimation process as initial values. The only parameter supplied by the user is the number of classes. All other parameters are estimated from the observed image. The algorithms have been tested on a variety of real color video sequence