Automatic segmentation of object region using Graph Cuts based on saliency maps and AdaBoost

In conventional methods for region segmentation of objects, the best segmentation results have been obtained by semi-automatic or interactive methods that require a small amount of user input. In this study, we propose a new technique for automatically obtaining segmentation of a flower region by using visual attention (saliency maps) as the prior probability in graph cuts. First, AdaBoost determines an approximate flower location using a rectangular window in order to learn the object and background color information using two Gaussian mixture models. We then extract visual attention using saliency maps of the image, and used them as a prior probability of the object model (spatial information). Bayes' theorem gives a posterior probability using the prior probability and the likelihood from GMMs, and the posterior probability is used as t-link cost in graph cuts, where no manual labeling of image regions is required. The effectiveness of our approach is confirmed by experiments of region segmentation on flower images.