Research On Graph-based Image Segmentation Method

Image segmentation is a basic task for image processing and computer vision which has used in medical imaging,security,and autonomous driving.The graph-based image segmentation model models images as a graph to represent pairwise correlations.This type of algorithm has great advantages in segmentation accuracy.But because of its poor efficiency and complex optimization,it cannot be used for real-time tasks.This paper finds and analyzes the problems in graph-based segmentation models,and proposes strategies to enhance their efficiency and segmentation accuracy.(1)This paper proposes a Dynamic Random Walk(DRW)model to enhance the efficiency of the random walk and solve its lack-seed problem.Our DRW uses dynamic nodes to cut off adjacent areas to reduce redundant walks and improve the efficiency.By weighting the random walk entropy to consider the similarity consideration between two non-seed nodes and use dynamic nodes to enrich the regional features.In addition,this paper proposes a 3D seed initialization strategy to apply the DRW into superpixel segmentation.Our DRW ensures boundary adherence and has much more higher efficiency than other graph-based models.(2)This paper also proposes an Exploring Normalized Cut(ENCut)model to solve the excessive normalization to enhance the small objects segmentation with high efficiency.The ENCut uses the meaning self-circulation graph model to promotes the degree of nodes in significant small objects which narrows the balance strength between the small and large regions;Then,the ENCut energy function is also used to diffuse the meaningful self-loop and reduces the cutting term of the salient region which can solve the excessive normalization problem.(3)This paper also proposes a Random Walk Refining Term(RWRT)to solve the over-global problem of the normalized cut and enhance the twigs segmentation.The RWRT uses pseudo-seed nodes and hierarchical graph models to make the random walk model unsupervised.It can be solved combined with the ENCut with the help of move-making algorithm.When combing with the RWRT,our ENCut not only reduces the redundant region but also enhances the small objects and twigs segmentation,which greatly improves the segmentation performance of the algorithm.