Study Of Thresholding Methods Based On The Image Spatial Co-occurrence Information

As one of the key method, thresholding technique is applied widely in image processing. As an important class of thresholding methods, more fully describing the image characteristics, applying the spatial information, the thresholding method based on the co-occurrence matrix obtains the stable and reliable segmentation results. In this paper, research is started from the classifical thresholding methods, and focused on the selection of feature information, the construction of co-occurrence matrix, generalized co-ocurrence of information. Several effective and feasible thresholding algorithms are improved.1.The application limitation of Otsu method is discussed, using image histogram information, the Otsu thresholding method focused on the object is proposed. Considering the degree of the homogeneity between the object and background, a thresholding algorithm based on the relative uniformity between the classes is proposed. Fusing the other information of gray distribution, an improved Otsu thresholding method is constructed based on generalized co-occurrence information. For the images with large uniformity difference between two classes, based on the gray level histogram weighted modified method, more reasonable threshold value is obtained. The independent weighted and the optimization selection of weight coefficient, improve the adaptability and the integrity of object. The improved method with neighborhood spatial mean and median information, without increasing the amount of computation, greatly improve the removal ability for noise.2.The gray transition symmetrical co-occurrence matrix thresholding is presented, for the deviation using the square distance method in classification estimation to the kew and heavy tailed distributions images, the region median values are defined of object and background, the median is used as the statistical value for image gray probability distribution information, the square distance-based symmetrical co-occurrence matrix thresholding based on the median is proposed, and the multi-thresholding algorithm is advanced. The relative homogeneous between object and background Otsu method is applied to the square distance symmetrical co-occurrence matrix method. In this way, a relative entropy-based symmetrical co-occurrence matrix thresholding method is derived, this method makes full use of the spatial information, obtains the more effective results, and is the basis of the parameters relative entropy thresholding methods base on the symmetrical co-occurrence matrix.3.Based on the “gray-gray mean” asymmetrical co-occurrence matrix method, for sparse histogram image, the fast algorithm of maximum entropy is modified, the 1D and 2D maximum entropy fast calculation formula are obtained, the operation speed is improved greatly. Analyzing the parameter optimization method of Arimoto generalized entropy, a expression error of existing 2D Arimoto entropy thresholding algorithm is proposed, a correct 2D Arimoto entropy method and a recursive formula are given, in case of the rich image edge and noise information, a better segmentation effect can be obtained.4.For the traditional “gray-mean” asymmetrical co-occurrence matrix, in view of poor adaptability in a single template, based on the minimum variance criterion, “gray-minimum variance mean” asymmetrical co-occurrence matrix is constructed. This method is sensitive to image edge information, for the images with rich boundary information, it has better segmentation effect, and works well in suppressing Gaussian noise. Based on the “gray-minimum variance mean” asymmetrical co-occurrence matrix, the relative entropy thresholding method and 2D Arimoto entropy Linear-type thresholding method are improved. The experimental results show that, the proposed method not only is notable in reducing the noise interference, and obtaining the more complete information.5.The application of the gradient information is studied in asymmetrical co-occurrence matrix thresholding methods. The “gray-gradient” maximum weighted conditional entropy formula is proposed. Applying with the uniformity and shape measure function, the weight coefficient is introduced for the different emphasis. The gradient information is applied to the thresholding method focusing on object, the problem caused by focusing only on the internal of the objects and backgrounds is improved. On the basis of the “gray-gradient” co-occurrence matrix, the mean information is introduced, the “gray-mean-gradient” 3D asymmetrical co-occurrence matrix is established, and a new 3D Otsu method is proposed, a fast recursive algorithm is presented. Compared with the “gray-mean-median” 3D Otsu method, this new method can obtain more complete and clearer edges, and effectively improve the ability of anti-Gaussian noise.