Seed growing for interactive image segmentation using SVM classification with geodesic distance

Abstract

In an interactive image segmentation, the quantity of a user-given seed is known to affect the segmentation accuracy. In this Letter, we propose a seed-growing method expanding the quantity of a seed to reduce the bias of the given seed and improve the segmentation accuracy. To grow the given seed, a supervised classification framework with geodesic distance features is proposed. From a single input image, a support vector machine (SVM) classifier is trained on the seed superpixels of an input image. Other non-seed superpixels are then classified into object, background and non-seed regions by the trained classifier. In experiments, the proposed method showed promising results by improving the segmentation accuracy of existing segmentation methods in public benchmark datasets.