Research On Image Segmentation Based On Super-pixel Fusion

As the basic step in the image process technology,image segmentation can divide images into regions and extract interested areas.These regions have similar characteristics.The result of image segmentation will directly affect the image analysis and image comprehension steps.The research of image segmentation technology has great significance.In the traditional segmentation algorithms,most of them regard the single pixel as the processing units,so these segmentation method have low computational efficiency and poor segmentation effect.Therefore,in order to reduce the computation cost,the method of dividing pixels by super-pixels has been developed.On the basis of super-pixel,this paper realizes the segmentation processing of image by two stages: pre segmentation and super-pixel fusion.The main research work is as follows:(1)Super-pixel generation algorithm is used for image pre processing.Optimizing the simple linear iterative clustering algorithm,by re-marking the super-pixel blocks,reduce the result of over-segmentation.(2)Updating the database of the segmentation algorithm based on region maximum similarity.Adding the super-pixel results that generated by SLIC algorithm into the database.In the original algorithm,the super-pixel results are obtained by mean shift clustering algorithm,and the shape of the super-pixel is irregular.The adjacent structure of the super-pixels has changed,which makes the segmented contour more standard.What`s more,when the super pixel feature is expressed,the color feature is simplified to improve the region merge speed.Experiments prove that the improved method is effective(3)In order to achieve a detailed division of the foreground and background of the image,a novel super-pixel fusion method is proposed.Regional relationship is represented by a region adjacency graph,the regions are taken place by nodes,the fusion process is judged by weights of edges.The energy function is set according to the color feature and the texture feature,and the energy function is used instead of the weight to fuse the two regions with the largest weight.The segmentation results are evaluated according to the evaluation criteria.The methods proposed in this paper are experimented with four othersegmentation methods in BSDS500 database.