Based On The Bayesian Level Set Image Segmentation Methods In Remote Sensing Image Building Extraction

The buildings in remote sensing images are of diversity and high complexity.The successful building recognition can provide theories and methods which are of universal guiding significance for other remote sensing image understanding.There are usually two strategies from monocular remote image:Edge-driven strategy and Region-driven strategy.The Edge-driven strategy could extract regular buildings only.The Region-driven strategy could extract relatively complicated buildings,but the extraction results depend on the segmentation results to a large extent.In this paper, we mainly study the remote sensing segmentation methods so that we can extract buildings better by using Region-driven strategy.Image segmentation is a basic step and key links in image processing.At present,a lot of segmentation methods have been proposed.The remote sensing image has its characteristics such as: large information,small object,diversified types and large noise.Therefore,the algorithms for remote sensing image segmentation are relatively less.In recent years,active contour model implements image segmentation by mathematical model.This model can integrate low-level visual feature(such as:edge,texture,gray value,color information) and experience knowledge(such as: prior shape,empirical statistics of color information) into this model,and provide a unified solution to a series of computer vision problems.Thus,this method is more extensively researched and applied in image segmentation field.The basic idea of active contour and segmentation is:firstly, giving one original contour(also called "curve") or multi-contour;secondly,defining an energy function about the image information and minimizing the function;thirdly,evolving the curve which will stop evolution at some characteristics of the image(such as image gradient).When we use the model to implement image segmentation,we should express our energy with our selected curve evolution way(parameters form or indirect form with level set method),and obtain a corresponding partial differential equation(PDE) by adopting the variational methods to minimize the energy function.Then we evolve the curve with discretized PDE iteratively.Active contour models can be divided primarily into parametric active contour(PAC) and geometric active contour(GAC).The PAC model expresses the curve deformation with parameters form directly so as to unsolve the topological change.But the GAC model expresses the curve deformation indirectly with level set function which always keeps effective and continuous so that it allows topological change during curve evolution.Therefore,we discuss the GAC model in this paper,and apply it to building extraction of complex remote sensing.In this paper,firstly,we introduce the basic theory of level set method and some classical GAC model.Secondly,we discuss the C-V model based on Mumford-Shah.Then,we develop a GAC model based on Bayesian theory.Finally,we propose an active contour image segmentation algorithm based on region competition which can extract multi-class object with fast evolving speed and high segmentation precision.The innovation of this paper can be listed as follows:(1) We analysis the C-V model which need re-initialization the level set function to a signed distance function(SDF) to maintain stable curve evolution and ensure desirable results.We improve the C-V model upon the shortcoming of re-initialization.The improved algorithm can speed up the curve evolution,and solve the problems such as when and how to re-initialize the level set function to a SDF.(2) We develop a GAC model based on Bayesian theory.Our model describes the C-V model from the perspective of mathematical statistics.It is a general region-based image segmentation model.In this model,we deduce a probability with Maximum a posteriori(MAP) according to the Bayesian theory,and define the energy function based on Bayesian MAP.Then we optimize the function by combining GAC model.In order to implement color image segmentation and make full use of the correlation between the color features,we adopt multivariate Gaussian density to represent the information of each region.In additional,it is easy to integrate other image information such as texture and shape into this model,and we can choose different probability function to describe various types of images.(3) We propose a fast image segmentation algorithm based on region competition which can incomplete multi-class object extraction.We adopt the region competition strategy to guarantee the partitioning unoverlapped during the curve evolution.This evolution strategy can realize arbitrary number class's segmentation.We achieve a new fast PDE,combining the existing PDE.Thus,we improve the speed of the curve evolution and ensure the segmentation precision at the same time. Our model can segment images into a fixed class-number but arbitrary shape of regions,and the same type object could be discontinuous.Furthermore,our model not only can incomplete color images segmentation,but also gray images.In addition,the energy function and curve evolution strategy are independent.