MDS-based segmentation model for the fusion of contour and texture cues in natural images

In this paper, we present an original image segmentation model based on a preliminary spatially adaptive non-linear data dimensionality reduction step integrating contour and texture cues. This new dimensionality reduction model aims at converting an input texture image into a noisy color image in order to greatly simplify its subsequent segmentation. In this latter de-texturing model, the (spatially adaptive) non-local constraints based on edge and contour cues allows us to efficiently regularize the reduced data (or the resulting de-textured color image) and to efficiently combine inhomogeneous region and edge based features in a data fusion/reduction model used as pre-processing step for a final segmentation task. In addition, a set of color/texture and edge-based adaptive spatial continuity constraints is imposed during the segmentation step. These improvements lead to an appealing and powerful two-step adaptive segmentation model, integrating contour and texture cues. Extensive experimental evaluation on the Berkeley image segmentation database demonstrates the efficiency of this hybrid segmentation model in terms of classification accuracy of pairwise pixels in the resulting segmentation map and in the precision–recall framework widespread used for evaluating contour detectors.