Font Size: a A A

Research On Image Segmentation Based On Fuzzy Theory And Region Methods

Posted on:2009-09-19Degree:DoctorType:Dissertation
Country:ChinaCandidate:S L LiuFull Text:PDF
GTID:1118360245479142Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Image segmentation is just to segment an image into different sub-images with different characters and get some interested objects. It plays an important role in image engineering, and is applied in a lot of fields such as computer vision, pattern recognition, medical images and so on.In most segmentation algorithms, methods based on fuzzy theory and region are very important and widely applicable. Comparability and uncertainty are ubiquitous between the object and background, because of various factors in the process of image-forming. Fuzzy set theory and the corresponding fuzzy information processing technique show great superiority in the process of describing and dealing with fuzzy events and imprecise knowledge. Segmentation methods based on region make use of the inside homogeneity, and realize regions' segmentation depending on the common attributes. The aim of segmentation is to endow every pixel with specifically attribute. So, in the paper the segmentation algorithms based on region and fuzzy theory are researched. In methodology, we make our research in three aspects: one is the information entropy based image segmentation technique, the other is the transition region extraction based image segmentation technique, last is graph theory clustering based technique. Then we do some comparison to each method' segmentation efficiency.Combining the fuzzy theory and probability theory, a new segmentation approach of 2D membership partition's Renyi entropy is proposed, then expand to 3D, ulteriorly. To an image, first, we do the fuzzy probability description. Then, conditional probability and conditional entropy are used to define fuzzy partition entropy. Experiments show in 2D histogram the segmented region includes a convex region constituted by a straight segment and high hyperbola. It is a crucial region to distinguish pixels between the object and background. The latter uses information of gray level distribution and neighborhood space. And the crucial region is composed by a plane and surface of high degree. Results show better stability and effectiveness.By analyzing the drawbacks of the traditional transition region extraction approaches, we know that the singlet characteristic would not be gained even little fluctuate showing in the background which leads great change in the curve, when we compute the W-EAG curve of high-cut and low-cut. Thus makes the reliability of transition region become poor. So, we give a modified transition region extraction and segmentation method based on fuzzy morphologic pretreatment and energy characteristic ratio transform. And more perfect gray distribution range can be gained.The traditional graph-theoretical clustering methods process the local area's joint characteristic of sample data as main information, which leads the most joint data in clusters can not be settled. Initializing each pixel as a class makes a large body of processing data. So we put forward a new graph-theoretical clustering algorithm based on spatial relation. The new one not only considers the relation between pixels and their neighbour hood regions, but also makes classes depending on uniform grey level of pixels. The required storage space and realized complexity get satisfied improvement to the traditional algorithms. Further, the maximum tress clustering method based on fuzzy graph-theoretical sets is proposed, when fuzzy similarity relationship is used to define edges' weight values while generating the complete graph. And we establish a more simple and applied fuzzy similarity relationship to describe the imprecision of combination regions.At last, we give a suit of segmentation evaluation system. The method uses fuzzy degree to measure segmentation quality. A new non-linear mapping function of changing from special field to fuzzy property field is proposed, which differs from the one used in image segmentation. Construct the model of evaluation and evaluate segmentation results of these segmentation methods by using many real images.
Keywords/Search Tags:Image Segmentation, Fuzzy Theory, Region method, Information Entropy, Transition Region, Graph-theoretical Clustering, Evaluation
PDF Full Text Request
Related items