Research On Multidimensional Otsu Thresholding In Image Segmentation

The ultimate goal of image segmentation is to segment the given image to meet therequirements of high quality target area. At present, there are nearly one thousand kinds of differentimage segmentation methods, Otsu algorithm is one of the most classical threshold segmentationmethods, which has developed from1-D or2-D to multidimensional field. Though it improves thequality of image segmentation with the increase the time complexity of algorithm, Otsu algorithm islimited in practical applications. This dissertation studies the multidimensional comprehensiveimproved algorithm of Otsu method based on the research of the present Otsu method.First of all, this paper analyzes the current situation of the development for Otsu algorithm,discusses the efficiency of different Otsu algorithms and then studies the various improvements ofOtsu algorithm in logic and dimensions. Finally, it leads to the comprehensive improvemedalgorithm of2-D Otsu method and3-D Otsu method to verify its feasibility and efficiency byexperiments. The study includes the following three parts:In the first part, this paper gives the comprehensive improved algorithm of2-D Otsu methodbased on the study of the existing2-D Otsu algorithm. Firstly, there is much redundant calculationof2-D Otsu algorithm and the statistical area is not reasonable enough. In order to solve thisproblem, one-dimensional threshold is used to replace two-dimensional threshold. The statisticalarea is selected to be belt type region on both sides of the diagonal line. Secondly, the recusivefunction is deducted from calculating the between-class variance to make recusive improvement.Thirdly, the weight coefficient is increased for pixel value and the neighborhood. From the abovethree points, it produces the comprehensive improved algorithm of2-D Otsu method. As a result,the time complexity descends fromO(L4)of original method toO (L2). The contrast experimentshows that this algorithm guarantees the segmentation effect as well as shortening the time.In the second part, this dissertation presents the study of the third dimension index selectionabout the3-D Otsu. Traditional3-D Otsu algorithm statistically adds a dimension in thetwo-dimensional basis; the key technology is to select the third dimension index. Some scholars putforward neighborhood median and weighted neighborhood median. But these indexes and seconddimensions (neighborhood average) represent the similar image information, so the effect is similarto2-D Otsu algorithm in experiment. Finally, various gradient pixel values are discussed, whichreflects the detailed image and edge information. So the Sobel operator is selected as the thirddimension index.In the third part, this paper puts forward the comprehensive improved algorithm of3-D Otsu method based on the study of the existing algorithms of3-D Otsu. Firstly, from the study of thethird dimension index selection, it identifies the three indicators which are pixel values,neighborhood average and Sobel operator. Secondly, three-dimensional threshold is replaced byone-dimensional threshold. The statistical area is changed to be the region around the threedimensional histogram body diagonal line. Thirdly, the recusive function is deducted fromcalculating the between-class variance to make recusive improvement, in order to exchange to thetime complexity from space complexity. Then, the weight coefficient is increased for the threeindicators because they reflect different degrees of image information. All of above, it produces thecomprehensive improved algorithm of3-D Otsu method. The time complexity descends fromO(L6)of original algorithm toO(L3). Besides, this algorithm guarantees the segmentation effectas well as shortening the time by contrast experiment. In the final analysis, it shows the using rangeof the two comprehensive improved algorithms.In a word, this dissertation studies various improved Otsu algorithm and puts forward thecomprehensive improved algorithm of2-D Otsu and3-D Otsu.