Research On Image Segmentation Based On Partial Differential And Variational Technique

With the continuous development of science in the fields of computer vision,artificial intelligence and thinking,digital image processing technology is moving towards a higher,deeper and broader level.Image segmentation,as a part of the technology,becomes a basic and key subject in the field of pattern recognition,object tracking,machine vision and image understanding,whose main purpose is to partition a digital image into multiple non-overlapping sub-regions with specific properties,and extract objects that match specific scenarios from background.Image segmentation has almost permeated through all aspects of image processing.So far,researchers have made great efforts to put forward many models and algorithms.Among them,the active contour model based on level set method allowing for topology change and information integration,has been continuously attracted the attention from researchers and scholars.However,due to the diversity,complexity and multiplicity of visual information,the image segmentation technology is still facing great challenges.Relevant researches are still being conducted at home and abroad in order to find a segmentation algorithm model with better universality and higher accuracy and efficiency.This is the significance of this thesis.Banking on the variational level set theory and the partial differential equation,this paper fully explores the classical active contour segmentation model based on the variational theory in order to overcome segmentation problems on images with complex background,intensity inhomogeneity and noise.The research has been carried out from perspectives of improving the sensitivity of models to the initial contour,its robustness against noise and its accuracy of segmentation,etc.And several segmentation models applied to noise pollution,intensity inhomogeneity and complex background have been proposed in this paper.The main contributions and innovative works of this paper are as follows:Firstly,to solve the segmentation problem against noise pollution and intensity inhomogeneous,a noisy image segmentation model based on local intensity difference is proposed.This model superimposed with a noise point repair function constructed on the local intensity difference,which enhances the robustness of the local level set model against various noises.So the noise point repair function can correct the noise image and the pixels deviating from the local intensity mean value.It can also adjust the pixel value within a reasonable range to reduce the interference of outliers to the energy functional calculation process and effectively reduce the impact of noise on the segmentation results.The corrected image reduces the measures of dispersion of the pixel value within the whole scale.It restrains the noise,improves the image quality which enables the image to be more smooth and effective,and meets the requirements of the segmentation technology to a significant extent.At the same time,a related energy model is established by this model via using the difference between the corrected pixel value and the local mean value.Experimental results on simulated artificial images superimposed with noise and natural images show that the proposed model is not only insensitive to the choice of initial curve position,but also has strong adaptability to the different types,and thus the intensities of noise and good segmentation results can be obtained.Secondly,traditional noise robust level set image segmentation methods are mostly devoted to accurately identifying the outliers and appropriate reduction the weight of the abnormal data to reduce their interference and influence on the segmentation process.Such methods discard so much normal image data that the segmentation accuracy of inhomogeneous and weak edge images in a noisy environment will be poor.In order to solve this problem,an automatic adaptive neighborhood active contour model is proposed in this paper.It uses the direction and length of the image gradient vector to guide the neighborhood deformation,that is to say,the neighborhood in the level set model is allowed to transform into an ellipse;the direction of the long axis of the ellipse coincides with that of the gradient vector at the center of the neighborhood;the eccentricity of the ellipse is positively correlated with the gradient vector.Under this deformation strategy,as the target boundary is clearer and the gradient value is larger,reducing the neighborhood scope and increasing the eccentricity along the gradient direction will discard some unrelated image information on the one hand;on the other hand,expanding the neighborhood and reducing the eccentricity will allow more image information to be involved in the calculation to ensure the effectiveness of the segmentation while increasing the precision of the segmentation.Furthermore,with the reference to the local correlation coefficient method and the degree of correlation between one point in the neighborhood and other pixels,this model downgrades the high noise and avoids the noise interference problem owning to abandoning too much image information in the process of neighborhood deformation.Multiple sets of quantitative experiments show that the model has a more superior performance than others in dealing with the Gaussian noise and weak boundary images.Thirdly,in view of existing local models,it can be found that the segmentation failure arises because they are so easily fall into the local minimum while segmenting images with intensity inhomogeneity that the evolution curve is halted in the background or inside the target and cannot continue to evolve.Therefore,we introduce the local intensity difference(LID)item into the RSF model and propose a local intensity difference model which can segment images with intensity inhomogeneity robustly and accurately.Since traditional local models only consider the difference between the fitted image and the original image while constructing the energy term,the grayscale variation of the image details is not comprehensively investigated.As a result,the curve will easily fall into the local minimum and stop at a wrong boundary point.The improved model further explores the deeper differences between the target and the background pixels and analyzes the interference mechanism of the wrong boundary points on the model.It makes use of the added local intensity difference items to comprehensively examine the pixels inside and outside the evolution curve in the neighborhood.In this way,it provides an accurate evaluation mechanism to exclude the wrong boundary points which cause the local minimum.In this model,the evolution curve is driven by the difference between the object and the background in the neighborhood while constructing energy terms and crosses the background or the object of the image until it is exactly at the object edge.A large number of comparison experiments on simulation images and real images show that the proposed model has a better segmentation performance compared with other local models.It can effectively solve the problem caused by the local minimum in local models.It shows greater segmentation accuracy and stability in segmenting different inhomogeneity images and has stronger robustness to the initial contour.