Image segmentation by cue selection and integration

In many recent works, image segmentation has been cast to a graph partitioning problem in which an affinity matrix represents the pairwise similarity of the nodes (pixels). In this paper, we develop an approach for the computation of the affinity matrix based on the combination of affinity matrices from various cues and its integration in the segmentation process. A principal components analysis (PCA) applied to the whole set of the normalized affinity matrices provides the uncorrelated relevant cues and their respective weights for the final combination. We then propose to integrate the evaluation of the affinity matrix at each iteration of an agglomerative algorithm in order to take into account the dynamics of the segmentation process. We finally define a criterion of satisfaction based on the variance–covariance matrix of the affinity matrices, which determines the end of the iterations. Experiments on a range of various images provide significant results.