Two-step Image Segmentation Based On Superpixel

Image segmentation is a basic topic in computer vision, and also is an important part of theimage analysis and processing. Image segmentation is the process of partitioning an image intomultiple disjoint segments(superpixels). The goal of segmentation is to simplify or change therepresentation of an image into something that is more easier to analyze and process. After thesegmentation, pixels in a region are similar with respect to some characteristics, such as color,texture or transparency and adjacent regions are significantly different with respect to the samecharacteristics. Although a large number of segmentation algorithms have been proposed overthe past few decades, we have yet to find one that could be applied to any image. At present,imagesegmentationhasbeenwidelyusedinmanyfieldsincludingtargettracking, humanfiguredetection, biomedical image and machine vision.Normalized cut is a graph-based algorithm for image segmentation. Normalized cut doeswell in image segmentation, but it is not suitable for real-time image processing for its timecomplexity.Superpixel is a meaningful atomic region which contain adjacent pixels that have somecharacteristics in common. Superpixels capture image redundancy, simplify the representationofanimage, andreducethecomplexityofsubsequentimageprocessingtasks. Hencetheycouldbeusedasapreprocessingstepforimageprocessing. Algorithmsforgeneratingsuperpixelscanbe categorized into graph-based methods and gradient ascent methods.Combining with the advantages of Ncut algorithm and Grid-based Simple Clustering al-gorithm, we propose a new superpixel-based algorithm for image segmentation. Firstly, wepre-partition an image into a set of superpixels using an over-segmentation algorithm calledGrid-basedSimpleClustering. Constructaweightedundirectedgraphusingcentersofthesuper-pixels. Then we partition the new graph using Ncut algorithm. Results of experiments showedthat the algorithm we proposed performs better in accuracy and efficiency.