Image segmentation via grabcut and linear multi-scale smoothing
Yi Lai, Chaohao Chen, Kun He · Proceedings of the 3rd International Conference on Communication and Information Processing · 2017
In order to improve image segmentation performance, a novel segmentation model is proposed combining GrabCut model and linear multi-scale smoothing. Multi-scale smoothing components, generated by Gaussian Kernel through an iterative scheme, provide different level image information that contributes to image segmentation. Each component is segmented by GrabCut model and segmentation result is different because the fine information is smeared step by step leading to that the appearance characteristic is different. A convergence condition, which rooted in the significant level of segmentation sub-regions on adjacent scale components, is constructed based on the invariance of object contour in each component. Compared to the traditional GrabCut, as the experiment result shows, the proposed model has a superior performance on real images and achieves better robustness against noise.