Interactive Image Segmentation using Graph Cuts

This paper presents an accurate interactive image segmentation tool using graph cuts and image properties. Graph cuts is a fast algorithm for performing binary segmentation, used to find the global optimum of a cost function based on the region and boundary properties of the image. The user marks certain pixels as background and foreground, and Gaussian mixture models (GMMs) for these classes are built using the colour and texture features of corresponding pixels. A likelihood ratio is used to calculate the relative probability of each pixel being foreground or background, based on the GMMs. Many features and their combinations are analyzed on images from the Berkeley Segmentation Dataset. Results of different segmentations are compared using precision-recall curves, F-score and accuracy. The average accuracy of the algorithm was 92% over a set of 20 images and the best accuracy was 99.5%.