The Study Of Algorithm Based On Cartoon-Texture Decomposition And Its Analysis In Image Segementation

In the recent years, with the development of computer science and technology, image segmentation technology has been the rapid development. Image segmentation is a technique and processes of that, a image is dived into several areas of particular and unique characteristic. It is a key step from the image processing to the image analysis, so it is widely used in the field of remote sensing, meteorology and military, and it has become the research object in computer science, information science, biology, medical and other disciplines. There are many widely used algorithms in image segmentation, such as image segmentation based on edge detection, image segmentation based on area growing and that based on threshold. However, due to the complexity and diversity of images, the researchers began to conjunction method in other fields, such as the segmentation based on lever-set, segmentation based on wavelet and that based on neural network. But there is not yet a common method used in various fields of the segmentation.Decomposition geometry provides a new method to describe the geometric characteristics of image, which image is decomposed into cartoon and texture is currently the most widely used field. The images that have decomposition characteristics often have the edges that are very irregular and complex features, their texture images are also complex and self-similar. Especially, the natural texture is very suitable for the decomposition model to describe. Because the image decomposition can remove the interference of texture, it has great significance for the image segmentation.Meyer model is the basic model of image decomposition; it has two functions F1, F2, wherein Fl is used to exclude the tiny shocks in the images while retaining smooth areas and strong edges (cartoon), and F2is used keep the image texture part. The major algorithms based on this model are Mumford-Shah model and ROF model. Because such models are nonlinear models, they are difficult to solve. So the linear methods have been taken to solve image decomposition problems, local total variation filtering algorithm is currently the best image decomposition method. However, whether it is a linear model or nonlinear model, they are all related to a decomposition parameter, and the selection of the parameter brings a big inconvenience to the process of image decomposition. In addition, for an image, the current algorithms only use one parameter, because the different areas have different characteristics, thus causes the effect that some areas have good results and some not.A new adaptive local total variation method based on credible data set is proposed in this article, which is based on local total variation and above shortcoming. The complexity of texture is different in different regions of image. The complexity of texture in the regions that contain small amount of texture but strong edges is low, in order to maintain the integrity of edges, we have to use a small parameter. However, in the regions that contain large numbers of texture, we have to use a large parameter to decomposition complexly. So this paper takes the idea of block processing and uses different parameters for different blocks to get a good result of the whole. In order to better describe the complexity of an image, we define a new function called oscillation according to the local total variation. The greater the oscillation is, the more complex the texture. Then we use the computer simulation method to construct the adaptive decomposition parametric curves based on oscillation. Thus, by calculating the oscillation of different images, it can get the decomposition parameter applied to the image, so the problem of selecting parameter can be solved. In addition, we use the integral image and the idea of the approximate Gaussian filtering to further reduce the computational complexity and dependence of decompose window to parameter. The experiments show that the results of new method are better, especially in specific application fields (medical image segmentation, tire defect detection, cloth texture detection and so on).