Research On Image Segmentation Method Based On Superpixel And Graph Theory

Image segmentation is a fundamental and important part in image processing,it has great influence on some more senior techniques in computer vision like target recognition and tracking,scene analysis,etc.Method of image segmentation based on graph theory is more popular because image can become a graph easily.In order to reduce nodes of a graph in image segmentation method based on graph theory and improve efficiency of image segmentation,superpixel is applied to image segmentation based on graph theory in this thesis.At the same time,we propose some improved algorithm.The main work of this thesis is listed as below:(1)A novel Flooding-Based Superpixel(FS)segmentation method is proposed to improve bad segmentation results of watershed superpixel segmentation algorithm while it only use gradient in flooding step.Firstly,our method uses color instead of gradient in flooding step,and the process order of each pixel is decided by color distance and space distance between pixel and its seed.In order to make boundary of superpixels near to image edge,we change boundary pixel iteratively by calculating the distance between pixel and its seed and the distance between pixel and seed of neighbor pixel with different label after flooding step.Compared to some famous superpixel segmentation methods,our method can get better segmentation result.(2)An unsupervised image segmentation method based on FS and spectral clustering algorithm is proposed to solve the problem of high complexity of spectral clustering segmentation method.Firstly,we use FS algorithm to get pre-segmentation result in order to reduce nodes of the graph,and feature fusion include color,covariance matrix,geodesic edge and space position of superpixel is used to calculate the similarity matrix.Experimental results show our method has higher segmentation accuracy.(3)Compared to unsupervised image segmentation,interactive segmentation method has more accurate result.So a novel interactive method based on FS and graph-cuts is proposed to reduce segmentation error rate.We use FS algorithm to pre-process firstly and a new energy function model based on superpixel and graph-cuts model is used,the region term of new model includes feature distance and geodesic distance between superpixels and labels.After getting initial segmentation result by using max flow min cut algorithm,we revise the region term of new model according to labels of feature nearest neighbor superpixels of every superpixel.Final segmentation result can be got by performing max flow min cut algorithm iteratively.Experimental results show our method has lower segmentation error rate.