The Improved Mumford-Shah Model And Its Applications In The Medical Image Processing

Segmentations of image is a basic but also difficult task in computer vision and image processing。The results decides the analysis, understanding, recognizes and also some other high-level module processing performance in image processing systems. Among these so many image segmentation approaches/algorithms, the Mumford-Shah (MS) model attracts a great attention because it can achieve image segmentation and image de-noising simultaneously. For this model always traps into local minimum solution, people have proposed many methods in order to find the global solution, one typical model is the Chan-Vese model (C-V model) which is proposed by Chan and Vese based on the MS model by introducing the level set method. This model didn't rely on the local gradient of image but use the global information of the image, and the segment results don't rely on the edge information, it also can avoid the effect of the noise, so it becomes very popular.Though the C-V model has those advantages mentioned above, it also has many disadvantages:(1) because the energy function of Mumford-shah model has an length term, this term corresponding to the mean curvature of equation in the numerical solution of the curve evolution equation, for some multi-object or high noise images, it costs a lot of CPU time, and is very sensitive to the parameters of the evolution equations, and it always receive local optimal results. (2) for those images that have big holes, it can't detect the structure inside the image steadly. (3) it cannot detect the edges of the image that have blurred edges accurately. In order to overcome the above shortcomings, we propose a fast improved hierarchical multiphase level-set segmentation model which based on the hierarchical multiphase segmentation model, from the expects of the mean curvature, level-set functionΦand Dirac functionδ(Φ) in the curve evolution equations and an introduction of an anisotropic diffusion equation, we modify the C-V model and get our improved hierarchical multiphase image segmentation model. Experimental results suggest that our model is more efficient and faster in segmentation of multi-object image and objects with weak boundaries.The C-V model also has many disadvantages:1. Lack of prior knowledge and structure information of the data model; 2. The segmentated results depend on the choice of initial conditions; 3. Unreasonable choice of weight factors will lead to bad segmentation results, if we only use experience to decide the values of the weight factors, it will lower the universality and auto processing ability of the algorithm. The Gaussian mixture model is a probability model which approachs the image histogram, also, it is a ideal model which can describe the slow varying of the gray level in the regions and the character of the whole image, so we introduce this model to describe the image,and use the poster probability to modify the C-V model. For we do this,we have three objectives:(1)no need to setting different parameter for the evolution equation for different images.(2) the single curve setting at any positions can segment the objects we are interested in.(3) we can segment the images accurately at the edge.C-V model also traps into local solution when it segment the images at the single scale. Multi-resolution methods obtain a global view of an image by examining it at various resolution levels, and perfectly unify the contradiction between the accuracy in the higher resolution and the easiness for segmentation in reduced resolution, then it avoids the local minimum. C-V model also lacks prior knowledge and structural information of the data model, but Markov random field (MRF) models provide a powerful and formal way to account for spatial dependencies between image pixels. So, we connect the multiresolution, MRF, and C-V model together, and propose a new hybrid model, and this new model can segment the multi-object images very well. Next, we introduce the denosing method which based on the Decimation-free directional filter banks, and we segment and denoise the image simultaneously at each resolution and get well results.