Image foreground extraction — A supervised framework based on region transfer

We propose a simple technique for extracting dominant foreground objects from images in this paper. Image foreground segmentation is a challenging task given problems due to complex background, clutter and occlusion, spectral variability in object appearance etc. Most of the techniques in this respect are semi-automatic in nature which require prior assumptions regarding the object location in the image in the form of rough bounding boxes. On the contrary, we propose a patch transfer based approach which propagates rough segmentation masks from the training to the test image patches. Further, these segmentation masks are combined and updated using a Markov random field model to generate the final foreground segmentation. The proposed technique is simple, robust and can work with very few labeled samples. Results obtained on the Weizmann Horse and the Leeds Butterfly dataset exhibit the superiority of the proposed approach with respect to existing methods from the literature.