Research On Multidimensional Otsu Thresholding In Image Segmentation

Image segmentation is the technique to separate images into regions which have different features and hence to ease the extraction of interesting objects. Image segmentation has long been a significant topic in fields such as image processing, pattern recognition, artificial intelligence, among others. It is also a key step in advancing image processing to image analysis, and plays an important role in image engineering. Image thresholding is one of the most important techniques in image segmentation. Its goal is to automatically find thresholds from the histograms of images, which can be used to perform image segmentation with the thresholds.Ostu thresholding, with its perfect performance, is one of the most popular image threshold segmentation methods. The algorithm can analyze image histograms automatically and obtain the best threshold value in statistics. It has been proved to achieve good segmentation results for images with objects in some background and can be extensively applied to the automatic segmenting of the objectives and background regions of image owing to its simplicity and the great adaptive capacity.In this thesis, several aspects of image segmentation are summarized, which involves the purpose and the significance, the present status in image engineering and the current research interests in segmentation techniques. Theoretical foundation in image processing and segmentation have been overviewed, including the basic concepts in digital image processing, definition, classification and evaluation in image segmentation, in which the theory and background knowledge is discussed in particular, and the present research progress of threshold segmentation has been investigated.In traditional Otsu thresholding, only variance between the objective and background classes is considered, but the classified information contained in every kind of pixels is ignored. This paper proposes an effective approach to improve Otsu thresholding, which combines cohesiveness of foreground and background image pixel class into standard Otsu thresholding and presents an improved classification function. Experimental results show that the improved method performs better in efficiency than the standard Otsu thresholding. Otsu thresholding in Multi-threshold Image Segmentation, as an important popularization of Ostu thresholding, has obtained accredited results. However, how to find the number of optimum threshold value still requires further research. In this thesis, a method of finding the number of optimum threshold value is proposed, and the relevant experiments and experimental result analysis has shown its applicability. With basic Otsu thresholding, the satisfied segmented result usually can't be obtained, because in the method, although information of grey levels of pixels in the image is utilized, the space relevant information among pixels is ignored to some extent, so the grey level histogram might not contain obvious crests and waves due to some influencing factors such as noise interferences, therefore some inaccurate segmentations might not be avoided. In consideration of analysis above, two-dimensional Otsu thresholding popularizes it on the two-dimensional histogram formed by grey level and average grey level of neighborhood on the basis of one-dimensional Otsu thresholding, and has improved the result of segmenting dramatically. But the segmentation effect of two-dimensional Otsu thresholding has dropped constantly with the increase of noise, hence the lowering of SNR. In recent years, the average value in the neighborhood has been introduced as the third characteristic and the histogram of three-dimensional grey level is structured, based on which the three-dimensional Otsu thresholding has been proposed, and the better result of segmenting has been obtained for the image with low contrast ratio and low SNR. This thesis made a further improvement on 3-D Otsu thresholding, in which the concept of value in the neighborhood weighted median is introduced, and thus the detailed information and edge features after segmentation can be better reserved. Since the introduction of three-dimensional space makes the calculating amount increase exponentially, this thesis proposes a fast recurrence formula in terms of mathematical deduction to reduce the complexity of calculation and the algorithm implemented has shown obvious performance improvement.