Exploiting Intensity Inhomogeneity to Extract Textured Objects from Natural Scenes

Extracting textured objects from natural scenes is a challenging task in computer vision. The main difficulties arise from the intrinsic randomness of natural textures and the high-semblance between the objects and the background. In this paper, we approach the extraction problem with a seeded region-growing framework that purely exploits the statistical properties of intensity inhomogeneity. The pixels in the interior of potential textured regions are first found as texture seeds in an unsupervised manner. The labels of the texture seeds are then propagated through their respective inhomogeneous neighborhoods, to eventually cover the different texture regions in the image. Extensive experiments on a large variety of natural images confirm that our framework is able to extract accurately the salient regions occupied by textured objects, without any complicated cue integration and specific priors about objects of interest.