We address the problem of object recognition in computer vision. We propose an invariant representation of the model and scene in the form of Attributed Relational Graph with focus on region based measurements rather than purely interest points. This approach enhances the stability of scene image representation in the presence of noise and significant scaling. Improved solution is achieved by employing a multiple region representation at each node of the ARG.The matching of scene and model ARGs is accomplished using probabilistic relaxation that has been modified to cope with multiple scene representation. The preliminary results obtained in experiments with real data are encouraging.