Researches On Level Set-based Active Contour Model Applied To Image Segmentation

Image segmentation is one of the most basic topics in digital image processing.This technology is widely used in medical diagnosis and analysis,remote sensing image positioning,automatic driving,assembly line product monitoring,sports competition,national defense military and many other fields.Active contour models which this paper focuses on are image segmentation methods based on strict mathematics,and their basic idea is to transform the image segmentation problem into the solution of functional extremum.Later,the introduction of level set method fundamentally changed the expression of curves and greatly promoted the development of active contour models.However,active contour models based on level set method still needs to be improved in some respects.For example,it is sensitive to the initial contour,easy to fall into local minimum value for complex images,and the evolution speed of level set is slow.In order to solve these problems,the main development of active contour models is systematically clarified,and then on the basis of previous work,a series of more perfect active contour models are proposed in this paper,which aims to complete image segmentation task more efficiently.The main research results are summarized as follows:1.The intensity distribution information of foreground or background region is introduced as the prior information of image segmentation and named by prior distribution.On the basis of nonparametric statistical active contour model,this paper proposes a distribution prior active contour model based on prior distribution information.In short,the prior distribution is embedded into the energy function of the existing model.Different from the commonly used prior shape information,the intensity distribution used as the prior information in the nonparametric statistical contour model greatly simplifies the construction and solution of the model.The idea of constructing energy function is to maximize the difference of intensity distribution in the interior and exterior regions of active contour curve,and minimize the difference of intensity distribution and corresponding prior distribution in each region during the evolution of level set function.The experimental results show that with the help of prior distribution information,the segmentation results are better for complex images.2.On the basis of the distribution prior active contour model,for the situation that the exact prior distribution information is unknown,this paper proposes a parametric expression method for the prior distribution by combining principal component analysis and level set method,and then proposes a parametric distribution prior model.The solution of the model is divided into two stages alternately.The first stage is to update the active contour curve using the variational level set method,and the second stage is to update the prior distribution in parametric form according to the current contour.The experimental results show that compared with the classical active contour model,the proposed model has a significant improvement in both segmentation accuracy and convergence rate.3.In this paper,a generalized divergence is introduced into the distribution prior active contour model to measure the difference between different distributions.The flexibility and robustness of the generalized divergence effectively enhance the segmentation performance of the model.In addition,a numerical scheme is proposed to determine the optimal divergence parameters by using gradient descent method.The experimental results of synthetic images and natural images show the effectiveness of the proposed model.4.In this paper,the shortcomings of the edge-based active contour models are analyzed,which is mainly sensitive to the position of the initial contour.Combining the global-to-local strategy and the idea of nonparametric statistical active contour model,this paper proposes a global-to-local region-based indicator,which can embed the region information into the edgebased active contour model.Different from the original edge-based indicator(which can not change the direction of the curve evolution),the proposed region-based indicator can allow bidirectional motion of the active contour curve,thus effectively making up for the defects of the existing models.