Interactive image segmentation based on hierarchical superpixels initialization and region merging

Abstract

As an efficient pre-processing technology, superpixel has been widely used in image segmentation, which improves the final results and reduces the complexity of the subsequent processing tasks. This paper presents an interactive image segmentation method based on the hierarchical superpixels initialization and region merging. It firstly proposes a novel hierarchical superpixels initialization framework including the coarse segmentation and the precise re-segmentation, which significantly reduces the number of superpixels, while keeping the boundaries adherence of superpixels satisfactory. Then the color histogram feature is applicable to the similarity measurement between two neighboring superpixels, and the Maximum A Posteriori Probability (MAP) estimation based on the color histogram is incorporated to estimate the probability that each non-marker superpixel belongs to the background. This paper establishes two different similarity measurements for the different stages in the MSRM merging procedure. Experiments verify the effectiveness of the proposed hierarchical framework to generate desired superpixels. The new interactive segmentation method can rapidly extract the object from the background.