Graph-based Superpixel Segmentation And Its Merging Algorithm

Superpixel segmentation is a hot topic in computer vision, using superpixels caneffectively reduce the redundancy of image local information, and greatly reduces thecomplexity and computational complexity of the image processing, but also retain theeffective information for further processing, the superpixel technology is more and morewidely used in many fields. This article explains a graph-based superpixel segmentationand its merging algorithm of image, we mainly research on several aspects: the edgedetection of image, the segmentation criterion of graph, the feature extraction ofsuperpixel regions, measuring the similarity between superpixel regions.Firstly, in consideration of the image boundary detection, we simply introduceseveral good boundary detection methods of image researched in recent years, we selectthe optimal boundary detection method through experiment contrast and analysis, theresult of this boundary detector is the input of the superpixel segmentation algorithm, thenwe introduces a kind of superpixel segmentation criterion based on graph theory, last wepropose a new superpixel segmentation algorithm which combines the edge detectionmethod with the segmentation algorithm based on graph theory, this algorithm make theimage segmentation algorithm which based on the graph theory can segment the images ofhigher pixel solution, this algorithm change the status quo which these previous superpixelsegmentation algorithm based on graph theory can only segment the images of lowresolution.Secondly, in terms of similarity measure, we put forward a new method of measuringthe similarity between two superpixel regions using the low level information of image,for example color feature, texture feature, position feature, shape feature. Image becomesseveral superpixel regions after being segmented by superpixel segmentation algorithm,we measure the similarity between superpixel region and its adjacent superpixel regionsaccording to the method of similarity measure.Finally, aimed at the over-segmentation phenomena in image segmentation, wefirstly calculate the similarity between superpixel regions based on the similarity measure method, then to judge whether they are merged according to the selected threshold value,Superpixel merging algorithms can effectively improve the over-segmentation phenomenawhich universally exist in the image segmentation algorithms, and break the rule that onlyuse the color and position information of the image to merge the regions.