Research On Image Segmentation Method Based On Variational Model

Image segmentation is a fundamental problem in image processing,and is the foundation of image based pattern recognition and image understanding.Image segmentation has been widely used in target detection,medical diagnosis,object tracking,intelligent transportation and other fields.Selective segmentation and global segmentation are two popular image segmentation tasks.The former one aims to separate the target regions or objects of interest from the background,while the later one aims to separate all objects from the background.In recent years,a large number of models and algorithms have been proposed.Among them,the variational models have been widely attended because of good mathematical explanation.In this thesis,we focus on designing variational models for selective image segmentation and global image segmentation.The specific achievements are as follows.1)A new variational image smoothing model and an active contour model are proposed for selective image segmentation.The proposed method consists of two stages.The first stage is a preprocessing step,which aims to smooth the given image,thus reducing influence of noise,or cluttered background on the segmentation.The second stage performs selective segmentation on the preprocessed image.For the first stage,a weighted variational smoothing model is proposed,which can preserve the edge of the image well and filter out noise and small scale details.For the second stage,a modified active contour model is designed to achieve selective segmentation.In the proposed active contour model,the global sparse gradient field is used to depict image edges,because it is more robust to noise than the traditional gradient operator.In addition,we use a set of marker points to construct the initial contour,which leads to more effective and accurate segmentation.Extensive experiments on real medical images show that,the proposed smoothing model can greatly facilitate the second stage,and the segmentation performance of the proposed method is significantly superior to relevant methods in terms of either visual assessment or quantitative evaluation.2)A thresholding based selective segmentation method is proposed.To facilitate threshold selection,we propose a new image smoothing model.The smoothing model consists of a regularization term and a fidelity term,where the regularization term is based on a trained plug and play denoising neural network,and the fidelity term introduces weights dependent on a set of marker points in the target region.The smoothing model can effectively smooth out other regions while protecting the target region.Extended experiments show that the new smoothing model effectively facilitate the threshold selection and the threshold segmentation,and the obtained segmentation results are significantly superior to that obtained by relevant selective segmentation methods in both visual and quantitative evaluation.3)Based on edge and region information,we present a global segmentation method,which can effectively segment images with strong noise and intensity inhomogeneity.First,an existing denoiser is applied to the original image to obtain a smooth image,hoping to reduce the influence of the noise and weak edges on the later edge detection.Then,an active contour model is given for global segmentation of the smoothed image.The model consists of three fidelity terms: one edge based and two region based.The edge based fidelity term is an edge stop function depending on the gradients of the smoothed image,forcing the active contour stop evolving at the boundary of the target.The global region-based fidelity term aims to minimized the discrepancy between the smoothed image and its local fitting image,while local region-based fidelity term considers pointwise image fitting.In order to ensure the smoothness of the level set function,Gaussian filter is used to regularize the level set function.Extended experiments show that the model has good robustness to noise and intensity inhomogeneity.Compared with relevant segmentation models,the proposed model also has significant advantages in segmentation accuracy.