Study On Segmentation Algorithms Of Complex Scene Image Based On Level Set Curve Evolution

In recent years,with the development of science and technology and the improvement of living standards,people need to acquire and produce a large number of digital,image,video and other data every moment in their daily life.Among them,image data has become one of the most popular expression methods in the society due to its intuitive and easy-to-understand characteristics and rich expression information.Therefore,various processing technologies for image information have become an indispensable part of human life.As an important part of the image processing process,image segmentation technology has also emerged as the times require.Among them,the level set method due to its flexibility is strong,has a strong mathematical theory support,easy to expand the advantages of widely used in all kinds of image segmentation.However,because the image data acquired in real life may have complex scenes such as inhomogeneous intensity,noise,blur and low contrast,and multiple targets,coupled with the traditional level set method's slow evolution speed and strong dependence on the initial contour,it poses certain challenges to the realization of accurate and efficient image segmentation.To improve the robustness of the level set method for complex scene image segmentation,such as inhomogeneous intensity,noise,blur and low contrast,and multiple targets,enhance the insensitivity of model to the initial contour,and maintain high segmentation efficiency,this thesis mainly does the following work:(1)A double-weighted signed pressure force(DWSPF)model is proposed,which can effectively segment heterogeneous scene images such as inhomogeneous intensity,noise,blur and low contrast,and multiple targets,and can guarantee high segmentation efficiency.In view of the problem that the existing segmentation method based SPF has poor segmentation accuracy for images with inhomogeneous intensity.This thesis firstly combines Legendre polynomials and global gray level information into the SPF to enhance the robustness of the model for segmentation of inhomogeneous intensity images,and uses a coefficient to weight the influence degree of Legendre term and global term.Then another weighted factor is introduced as the coefficient of the gray level information fitting center of the inner and outer regions of the evolution curve to optimize the evolution of the curve to the target branches with complex shapes in the region of interest.Experimental results show that the proposed method has a good segmentation effect on heterogeneous scenes such as inhomogeneous intensity,noise,blur and low contrast,and multiple targets.Meanwhile,the model can effectively control the time consumption and enhance the insensitivity to the initial contour.(2)An adaptively weighted signed pressure force(AWSPF)image segmentation model based adaptive weighted SPF is proposed.Aiming at the time-consuming and inaccurate problem of manual setting of gray level information in the inner and outer regions of the curve to fit the center coefficient of the DWSPF model,and driving the evolution curve stay at the target edge better,the AWSPF model first defines an adaptive global average intensity(GAI)term based the image information of the inner and outer regions of the evolution curve.Then,an adaptive Legendre polynomial intensity(LPI)is defined based the image information of the inner and outer regions of the evolution curve.Finally,GAI and LPI are introduced into the new SPF at the same time,and their influence is weighted by coefficients.A new edge stopping function(ESF)is defined,which includes the region information and edge information into the gradient flow equation.The experimental results show that this model has the basic segmentation performance of DWSPF model and can also realize the adaptive weighting coefficient of the fitting center,and the curve can stay at the edge of the region of interest better due to the introduction of edge information.(3)A two-stage automatic segmentation system for application scenarios based faster region convolutional neural networks(Faster R-CNN)and AWSPF model is proposed.Since the level set method is a semi-automatic segmentation method,the initial contour needs to be determined manually to carry out iterative evolution,but the initialization process is quite time-consuming,and the position of the initial contour has a great impact on the segmentation results.On the other hand,because of the typical representation of medical liver images,liver segmentation was selected as a specific application scenario for this experiment.In this thesis,the liver image data from Chongqing Cancer Hospital are firstly expanded and manually annotated,and then the data is sorted into a standard data set.Then Faster R-CNN is used for model training and preliminary detection of the test images.The central point of the detected rectangular box is automatically located to the initial contour of the suspected liver region,saving the time cost of manual initialization.Finally,the AWSPF model based the initial contour is used to achieve the accurate segmentation in the second step.Experiments show that this method can achieve satisfactory segmentation results and avoid the tedious and time-consuming manual initialization steps.