Study On Image Segmentation Based On Saliency-Driven Regional Level Set Model

Image segmentation has been a hot and difficult problem in the field of image analysis.It is also one of the most important steps in image processing,which is a prerequisite for many higher-level image analysis,recognition,and understanding techniques.Generally speaking,image segmentation divides an image into a number of non-overlapping regions based on visual features such as grayscale or color values,edges,shape,or texture.These features then should exhibit uniformity or consistency within the same region,while being distinct from the other regions.In recent years,level set based image segmentation methods have become a representative and important method for high-precision image segmentation due to their topological theoretical basis and advantages in extracting regional grayscale information.Thus,it is necessary to study them deeply.However,level set image segmentation is still in the developing stage.The investigation of its theory and applications needs to be enhanced and improved.In this dissertation,we mainly focuses on level set image segmentation energy functional model construction and its practical applications.We study its theoretical properties from three aspects,namely the construction of saliency-driven energy functional,the design of energy functional similarity measures,and the construction of distance regularization energy functional terms.We also propose corresponding effective numerical solutions.The main work and innovations of this dissertation are summarized as follows:1.A novel saliency-driven local Chan Vese(LCV)image segmentation model based on fractional order difference images(FLCVSR)is proposed in this dissertation.In view of the shortcomings of most existing level set based segmentation models that need a manually setting of the initial contour and have distance regularization terms that cannot guarantee the stability of evolution well,we propose to use the spectral residual method(SR)to obtain the saliency map of a given image and the initial level set.Furthermore,we use the absolute value instead of the square root operation to simulate the fractional gradient of the image.With this,we can obtain the fractional difference image,and construct the distance regularization term based on the logarithm and polynomial function.By fusing the fractional gradient image into the LCV model,we finally construct the saliency-driven fractional order regional level set energy functional.Experimental results demonstrate that the FLCVSR model can obtain better segmentation performance on images exhibiting intensity inhomogeneity and noise as well as images with real-world scenes.2.A new automatic regional local salient fitting local Chan Vese image segmentation model(ALSLCV)based on local region salient fitting is presented in this dissertation.Existing regional level set models use local average information to simulate local regions.The segmentation performance of these models is susceptible to the influence of the initial contour and images with intensity inhomogeneity.Differing from the above methods,we apply the saliency detection method to obtain the initial contour,which places the initial level set near the target objects.We calculate the degree of change of the local area based on the median value of the original image,and obtain a new transformation by extracting the distribution information of the gray level change.With the idea of fitting the regional distribution information,we construct the salient local fitting term based on the transformed modality and further design a novel distance regularization term based on logarithmic functions and power functions.With this,we can avoid the need for reinitialization and improve the stability of the evolution.Experiments on images with intensity inhomogeneity and noise as well as images with real-world scenes confirm the high accuracy and robustness of our ALSLCV model.3.We design a sweeping optimization algorithm for the local Chan Vese(LCV)model and propose the saliency-driven local Chan Vese(LCV)model based on the cosine measure and its two variant algorithms.After the evolution equations are obtained by the level set method,they are usually solved numerically using the finite difference method,which is a simple and easy to understand algorithm but requires a large number of calculations and thus is too slow for many special applications.For this problem,we first construct a sweeping optimization algorithm for the LCV model and apply the cosine function similarity measure to represent the data fitting terms in the region-based level set image model.Then,we propose a quadratic polynomial-based distance regularization potential function.We further design the initial contours based on saliency detection,and present finite difference and sweeping optimization algorithms.The sweeping optimization algorithm avoids solving partial differential equations.It is robust to initial contours,and has the advantage of automatic evolutionary termination.The experimental results show that the proposed sweeping optimization algorithm has good performance.4.We develop a regional level set energy functional similarity measure framework.Level set image segmentation models have been a hot research direction in the field of image segmentation.Hundreds of region-based level set image segmentation energy functionals have been constructed,but none of them can deliver good segment performance over a wide range of different images,and too many model choices inconvenience the users.Based on three region-based level set energy functional proposed in this dissertation and the popular region-based level set energy functional models,we defined six categories of region-based level set segmentation energy functional general frameworks by combining energy functional similarity measures.Each category contains a variety of popular region-based level set models.Finally,the segmentation performance of different energy functional similarity measure frameworks is compared through experimental analysis,and their advantages and disadvantages are given.5.A survey of the popular distance regularization potential function is given and two new categories of distance regularization terms are proposed.In order to obtain accurate segmentation results,the level set method needs to be reinitialized periodically during the evolution process so that it always maintains as a signed distance function(SDF)near the zero level set.Various distance regularization terms have been developed to avoid reinitialization.In order to study the specific effect of each distance regularization term,three regularization terms from this dissertation and the popular distance regularization potential functions are analyzed,and the potential functions and their diffusion ratio functions are classified into five categories according to the different types of potential functions.We further propose a type of potential function based on logarithmic function and polynomials and a category of diffusion ratio function based on rational power function.Through comparative analysis,both types of distance regularization terms guarantee the signed distance function characteristics of the zero level set and achieve the validity and stability of the evolution of the level set function.