The Research On Image Segmentation Algorithm Based On Hierarchical Merge Tree

Digital image is an important medium for people to understand the world in the development of science and technology.Similar to human vision,computers can obtain and transmit information through digital images captured by devices to accomplish specific tasks.As the basis of digital image processing technology and key technology,image segmentation has been widely research,and there are many methods or models are applied to image segmentation.Among the image segmentation modeling method,the hierarchical merge tree which is a combination of region merging technology and the structure of the binary tree,has obvious advantages on providing more granularity and flexible rules embedded image description.However,the existing image segmentation algorithms based on hierarchical merge tree only consider single constraint embedding,which cannot meet the needs of complex applications such as traffic target detection and remote sensing ground object detection.Based on the hierarchical merging tree model,we make researches on the constraint embedded,and propose more effective and feasible method for multiple constraints embedding.Then for the new merge tree constructed from multiple constraints,we propose the evolution function analysis based on node attribute change to find the optimal tree cut and meet the application requirements of segmentation results.In order to improve the accuracy of segmentation results and verify the proposed method,we propose two new superpixel algorithm to provide reasonable initial status for constructing hierarchical merge tree on RGB-D images and hyperspectral images,and verify the effect of the proposed method in this paper in the experiments.The main contributions of this paper are as follows:1)This paper proposes the construction method of multiple constrain embedding hierarchical merge tree based on Bayesian theory.By exploring the structure of the tree and the characteristics and application of Bayesian theory,the region merge in the process of constructing hierarchical merge tree is transformed into the maximum posterior probability problem.In order to solve the problem of measurement inconsistency,redundancy and conflict among multiple constraints,the rules for describing the regional merge likelihood function and priori function are proposed in combination with the characteristics of tree structure,which makes it possible to construct a more reasonable hierarchical merge tree with multiple constraints.2)This paper proposes a cutting method of multiple constraints embedding hierarchical merge tree.In order to find the optimal cut of the tree to get the final segmentation result,and make the segmentation result conform to the priori of the constraints,the evolution function is constructed by exploring the node attribute changes along the tree path,and the change trend of the evolution function is analyzed to find the optimal hierarchical merge tree cut.3)This paper expands and applies the proposed multiple constraint embedding hierarchical merge tree based image segmentation method for RGB-D images and hyperspectral images.First,through studying the characteristics of the image and existing problems of superpixel segmentation algorithm,we propose deep learning based superpixel segmentation algorithm for RGB-D image and watershed transformation based superpixel segmentation algorithm for hyperspectral images,aiming at providing a shape rules and efficient initialization region for hierarchical merge tree.Finally,based on the two proposed superpixel segmentation algorithms,the effectiveness of the proposed method is verified in RGB-D images and hyperspectral images.