Automatic Segmentation of Objects of Interest from an Image

In this paper, we propose a novel variational energy formulation for image segmentation. Traditional variational energy formulation for image segmentation like that in [41, 30] only incorporates local region potentials. We introduce a global image data likelihood potential to address the problem that each local region usually contains a portion of incorrectly classified pixels during the iterative energy minimization process. By combining it with local region potentials, we obtain more robust and accurate estimation of the foreground and background distributions. The minimization of the proposed local-global energy is achieved in two steps: the evolution of the foreground and background boundary curve by level set; and the robust estimation of the foreground and background model by fixed-point iteration, called quasi-semi-supervised EM, which is particularly suited for the learning problem where some unknown portion of the data are labeled incorrectly. Extensive experimental results including business card extraction, road sign extraction and general object-of-interest segmentation, demonstrate the robustness, effectiveness, and efficiency of the proposed approach.