Content-based Superpixel Merging And Its Application In Image Segmentation

As the first step of many image analysis and recognition tasks, imagesegmentation is a very basic but important part, which greatly influences thefinal analysis results. Traditional image segmentation methods are mostlybased on pixels, i.e., they describe each pixel’s color, texture and gradientstatistics characteristic information. While these local characteristics lack ofinformation of neighborhood’s synergy statistics, which often makes thesegmentation results too fragmentized. Superpixel, as a widely used imageanalysis method in recent years, performs better in describing the regionalinformation, and reflects better robustness in processing the images withnoise, occlusion and shadow.This paper studied and summarized several current popular superpixelcalculation methods. Then, we researched the content expression andsimilarity measurement of superpixels. On the basis of the work mentionedabove, we applied the progressive merging algorithm and affinitypropagation clustering to over-segmented images, and eventually split themain objects from the background.In particular, the proposed superpixel progressive merging algorithm isbased on the basic scheme of graph theory: an image is mapped into anundirected weighted graph, where the superpixels are treated as nodes of thegraph, and edges link adjacent superpixels together, edge weight between twonodes are set proportional to the similarity between the superpixels. Then, wemerge the two superpixels with the smallest weight. Through the experiments’comparative analysis, our proposed superpixel progressive merging methodsperformed good segmentation results. Especially in regional segmentation offew categories, this method showed better competitiveness, laying a good foundation for further image semantic understanding and object recognition.