Saliency-SVM: An automatic approach for image segmentation

摘要

Although there are some support vector machine (SVM) based methods for image segmentation, automatically and accurately segmenting objects that appeal to human perception is indeed a significant issue. One problem with these methods may be that the human visual attention is seldom taken into consideration. This paper proposes a novel visual saliency based SVM approach for automatic training data selection and object segmentation, namely Saliency-SVM. Firstly, a trimap of the given image is generated according to the saliency map in order to estimate the prominent object locations. Then, positive (salient object) and negative (background) training sets are automatically selected through histogram analysis on trimap for SVM training. Finally, the whole salient object is segmented using the trained SVM classifier. Experiment results on a benchmark dataset demonstrate the effectiveness of the proposed approach.