A Foreground Extraction Approach Using Convolutional Neural Network with Graph Cut

Abstract

In the literature of computer vision, many techniques have demonstrated their potentials for interactive image segmentation. However, most of these state-of-the-art algorithms are unable to produce accurate boundaries without more user interaction, as they are highly sensitive to the seed's quantity and quality. These techniques frequently depend on more user interaction to refine the boundaries. In order to solve the problem and get accurate boundaries via less user interaction, in this work, a robust interactive image segmentation method is proposed based on generic multiscale oriented contours via single forward pass Convolutional Neural Networks (CNNs) and graph cut framework. We first utilize CNN to construct the boundary-level information and then combine this boundary-level information with the boundary energy term of graph cut framework. The proposed method exhibits a significant leap in robustness to user interaction, smooth boundaries, accurate segmentation and the ability to handle the changes. We show that the sensitivity to the seeds can be controlled and accuracy can be improved via boundary-level information. We further conduct both qualitative and quantitative experiments on benchmark datasets, showing that our proposed method outperforms the state-of-the-art interactive image segmentation techniques.