Stochastic Active Contour Model For Image Segmentation And Uncertainty Analysis

As a basic technology of image engineering,image segmentation plays a very important role in various fields,e.g.,pattern recognition,computer vision and artificial intelligence.Image segmentation,a process of classifying image pixels into different regions corresponding to the actual target,completes the separation of the target and the background,realizes the conversion of low-level data to high-level information,and makes high-level image analysis and image understanding possible.In the past few decades,thousands of image segmentation methods based on different theories have been proposed to realize this classification task.Among then,active contour models are known for the capability of accurate boundaries tracking.These methods drive a closed planar curve deforming and finally approaching to object boundaries based on a partial differential equation deduced from an energy functional.This energy functional could effectively combines low-level image data and high-level semantic information or priors,which greatly contributes to its popularity.After decades of development,the active contour model has gradually become an important branch of image segmentation,and has formed a series of relatively complete and systematic methodologies.However,partial differential equations employed in these methods to drive curve evolution could not well handle the uncertainty from data and model.In addition,most of the current methods lack uncertainty measurement for segmentation results,which makes against obtaining stable segmentation or forming a more accurate segmentation results.Considering that,we focused on the uncertainty in image segmentation,employed intuitionistic fuzzy set,stochastic partial differential equations,and Bayesian random parameter estimation method,etc.to empower active contour models to generate more stable and accurate results.The associated contribution of our study are as follows.(1)Aiming at the problem of the inability to describe the uncertainty of the data and the model,and the lack of uncertainty measures in the segmentation results,a new stochastic active contour model based on intuitionistic fuzzy set was proposed to segment images stably as well as to measure uncertainty of segmentation results.Our model firstly generates random images based on camera noise model and atmospheric scattering model to simulate the uncertainty from data,and then employs intuitionistic fuzzy set to simulate the uncertainty of segmentation model.Meanwhile,Brownian motion in stochastic partial differential equation helps easily crossing saddle points,and contributes to obtaining global local minimum of optimization.In addition,our model defines uncertainty by multiplying the variance of the segmentation results on stochastic image with the hesitation degree of the intuitionistic fuzzy set,and provides a piecewise measurement of uncertainty in segmentation.(2)Aiming at the uncertainty of the inability to describe the parameters and the timeconsuming and labor-consuming problem of manually adjusting the parameters,a new stochastic kernel active contour model was proposed to facilitate and optimize the parameter choosing of stochastic partial differential equation as well as to segment images.Firstly,the model obtains the posterior distribution of the random parameter vectors by using Bayesian random parameter estimation method.Then,the steady-state distribution and specific values of random parameter vectors are further obtained by Markov chain Monte Carlo method.The segmentation parameters could be trained only given the corresponding distribution range of random parameters,by which manual parameter adjustment is avoided.Finally,a stochastic kernel active contour model whose parameters are adjusted automatically is utilized to improve the segmentation performance on inhomogeneity.In addition,the variance decomposition method is used to calculate the sensitivity of the relevant parameters,which can be used to guide the image segmentation to optimize the selection of parameters.In summary,this thesis studied the uncertainty from data,model and parameters in PDE and proposed two stochastic active contour models for image segmentation.Besides realizing image segmentation,our works could provide the uncertainty assessment of segmentation results and sensitivity analysis of random parameters,which all contributes to realizing a more stable segmentation,avoiding manual adjustment of parameters and optimizing selection of parameters.Our thesis provides new thoughts and views to theories of active contour models and its application.