Study On Approach Of Image Segmentation Based On Decomposable Markov Networks

So for, there are not theories and methods in image segmentation. All kinds of algorithm on segmentation in existence almost dependents some specific issues, which is difficult in this domain. The paper poses a new theory and approach based on which we can improve the dependences of segmentation to concrete cases. So the method submitted to the paper, in methodology, have commonalities in a manner.Image segmentation is very important research content in image information processing. And it is far and wide applied in image understanding, pattern recognition, image coding and image composition, and so on. But it is classic trouble problem. Especially the edge between regions of image is hard to avoid bring about gray scale smooth due to mechanism of capturing image, and the variant object regions in image exist overlap of gray scale by reason of complicate action of light. These all engender difficulty of image segmentation. The effective averting method is pre-treatment of smoothing image, and it can widely reduce overlap of gray scale. On the other hand, the noise also can severely affect quality of image segmentation, so pre-treatment of image noise filter is important aspect.The method of decomposable Markov networks (DMN) posed by the paper is based on principle of cross entropy of system to conduct searching structure of networks, and the brightness of pixel is judged by the output of networks structure, and classified into three sorts: high bright, in-between bright and low bright, so buffer region be get. So-called buffer region just is a set in which its pixel brightness can't be judged by local evidence. It is different from transition region, but there are connections between them. Namely, the buffer region comprises transition region, and transition region is not surely buffer region, because they have unequal idea. Buffer region is container of pixels whose brightness can't be judged temporarily. We can control quantity of pixels in buffer region by modification of capacity of buffer region. The pixels escaping from buffer region self-adapting may be respectively belonged to image region of high or low brightness, thus image segmentation can be carried out by control method of buffer region. The process is continuous. But transition region is regarded as a part containing image region' edge, and segmentation of it can get edge between regions, namely transition region is fixed, and no-adjusting one. So the result of image segmentation may be hard to avoid to be affected by noises and errors.In this case that DMN (decomposable Markov networks) model can judge belongingness of every pixel brightness, so it can also carry out image smooth and filter. Materially, no matter image smooth, filter or segmentation, its key is recognition of single pixel characteristics. Since DMN can identify characteristics of single pixel, it certainly can put into effect image smooth, filter and segmentation based on one same model. Some correlative literatures all hold that image smooth and filter profit to image segmentation, so the method of image smooth and filter is the paper's study contents. The image smooth, filter and segmentation methods submitted is more and more, but they almost are methods based on concrete image characteristics, and they can't smooth, filter and segmentation image by one kind of model.In image smooth or filter process, the no gray quivering or no noise pixels should be unchanged after and before smooth or filter, and the operation of smooth or filter should only change the value of gray quivering or noise pixels. For waiting for smooth or filter, the three important information of pixels can be get: first, the precise position of gray quivering or noise pixels; second, the tendency of adjustment of gray scale value of gray quivering or noise pixels; third, the precise position of near neighbor pixels with maximum correlation with gray quivering or noise pixels.The traditional method of smooth and filter (inclusion of mid-value and mean) have mortal defectiveness. Namely, first without precise position of gray quivering or noise pixels, so the methods consider that central pixel in window is pixel of noise; second the methods consider that all pixels in window effect for pixel of noise. So when smooth and filter, the substitute for noise pixel is mean value of all pixels in window without distinction. Thus, because the traditional methods of smooth and filter (inclusion of mid-value and mean) do not locate precise position of gray quivering or noise pixels, it gives rise to distortion of geometry of image. And because the traditional mean (or mid-value) methods do not distinct correlation between central noise and near neighbor pixel, it gives rise to distortion (blurring) of gray scale of image. The traditional methods almost exist more and more serious distortion of geometry and distortion (blurring) of gray scale of image with larger filter window, and this is a pair of contradiction without harmonic.The characteristics of method of image smooth and filter based on DMN is that first precise position of gray quivering or noise pixels is located, then concluded the precise position of near neighbor pixels with maximum correlation with gray quivering or noise pixels, finally only change the value of gray quivering or noise pixels by mean (or mid-value) near neighbor pixels correlation with gray quivering or noise pixels. Obviously, the precise noise position ensure stability of geometry of image, and selectivity mean (or mid-value) computing ensure stability of gray scale of image. So the method synthetically ensures quality of image smooth and filter, and put pillar stone of improving image segmentation.For carrying out above principle, the paper conducts extensive correlation research.The paper studies closely the intrinsic connection between graph theory and probability network, and the essential characteristics in decomposable Markov networks (DMN). In the meantime, the applications, developments and limitations of graph technology in image segmentation are treated quite overall. A method of extension and building model of DMN in image segmentation is brought forward. Finally, some experiments are implemented in order to analyses and compare the difference between the various other methods and methods by paper.It is pointed that the method we posed has widespread suitability in basic image processing field. Because it not only can segment image, but also smooth and filter image, if its rule in output is changed. Generally, not only the smooth and filter of image is regard as independent study domain, but also they are a preprocessing mode before image segmentation. So, the study on them is either by-product or important supplement for image segmentation method posed by our paper. In addition, in imagesegmentation method based on DMN, we submit some new concept such as buffer region, double image system and so on. And the paper analyses, and compares their difference with transition region extraction method (TREM) and morphological Watersheds (M-WS) image segmentation. The conclusion states clearly that the study of our paper has extensive theory worth and practical meaning.The major contents of study include:?The difficulties and present situations of image segmentation systematically are summarized. It is pointed that image segmentation is very important in image understanding, artificial intelligence and computer vision. And some of concrete algorithms based on neural network and information theory are analyzed, and expounded.?The significance and method of probability network (PN) is systematically probed into which include background of its creation and present situation of development. And the profound relationships among graph, neural network and probability network are revealed. The extensive PN implications, and its intrinsic relations with MRF and neural network all opened up our thinking in domain of image segmentation.?The technologies of image segmentation based on graph are deeply surveyed. And the limitations of them only based on graph are apropos discussed. And it is pointed out that the graphic construction and output based on graph should induct various mathematics theories and methods such as mathematical statistics and information theory, in order to construct more perfecting model of graph and output methods based on it.?A new definition of DMN used for image segmentation is drew, and the ways extended are researched.?The model of DMN for image segmentation is established, and analyzed. And the relevant issues such as image smooth and filter are investigated, and inquired into.0 Some experiments based on above-described theory and approach is implemented, and investigated. And the corresponding conclusions are obtained.?The overall summary and analysis about MRF for image segmentation is conducted. And the relationships and differences between MRF and DMN have in-depth probe, and the conclusions are reached by comparison with tow methods.The major innovations of study include:?The relations between graph and probability network are summarily studied poignantly.0 The paper posed the model and method based on DMN for image segmentation.?The principle and method between in MRF and in DMN for image segmentation is compared.?Our DMN model can be used main tasks of image processing such as image smooth, filter and segmentation. And it is studied by experiments.?Some new concepts such as buffer region is created.The problem in existence:?There are few errors in edge of image for segmentation noise image due to theprocess of smooth or filter.?The parameters based on DMN method such as radius of hunting networks and factors of buffer region must be obtained by experience, so the automation of our algorithm is limited.The future study tasks required:?The image filter and smooth should be fit into organic process of noise image segmentation based on DMN. And namely, the image filter, smooth and segmentation based on DMN may be integrated to fulfill in order to reduce edge errors.?The intrinsic relations between image segmentation and the parameters such as radius of hunting networks and factors of buffer region should be established with the help of other theories such as spatial statistics and variation analysis. Then the automation of our algorithm may be improved, and the dependency of experience also may be cast off.?The characters of buffer region wait for in-depth study.