Efficient image labeling based on Markov Random Field and error backpropagation network

Abstract

A Markov Random Field (MRF) model-based approach is proposed as a systematic way for modeling, encoding and applying scene knowledge to the image understanding problem. First, the image is segmented into a set of disjoint regions and a Region Adjacency Graph (RAG) is then constructed from the resulting segmented regions based on the spatial adjacencies between regions. The problem is then formulated by defining region labels and these labels are modeled as an MRF on the corresponding RAG. The knowledge about the scene is incorporated into an energy function that consists of appropriate clique functions which constrain the possible labels for regions. However, in the image interpretation problem, it is difficult to find appropriate parameter values of the clique functions since the real scenes are variable from image to image. The clique functions are implemented as error backpropagation networks so that they can be learned from sample training data. The optimal labeling results are then achieved by finding a labeling configuration which minimizes the energy function through simulated annealing. As preliminary experiments, the proposed method is exploited to interpret the color scenes.