Region Segmentation Methods Based On Histogram-Independent Rule And Based On Combining Control Strategy

Image segmentation is a key step in image understanding, and it is also a classic problem in computer vision. Thresholding segmentation and watershed segmentation are two kinds of region-based image segmentation technologies. Many existing thresholding methods can only deal with the images with certain histogram pattern, which limits their application areas. The main problem with watershed segmentation is its inherent over-segmentation. Image enhancement and region merging is two common strategies for suppressing the over-segmentation. However, for some structure texture images in biomedical engineering field, the segmentation result of watershed transformation accompanying with image enhancement or region merging strategy will still result in over-segmentation or under-segmentation. In view of these cases, one of main aim is to develop several new thresholding methods with the histogram-independent capability. Another aim is to develop more robust watershed segmentation method for segmenting accurately the structure texture images.Many existing thresholding methods can only deal with the images with certain histogram pattern. For overcoming this problem, three kinds of novel thresholding methods are proposed, and all of them are independent of the gray level histogram of the images. Firstly, we develop a novel image bi-level thresholding method based on multiscale gradient multiplication (MGM). We analyze the advantages of MGM on the location accuracy and anti-noise, and we deduce the reasonable values of the filter time and the filter scale of MGM. Through analyzing the statistical distinction of MGM between the edge and non-edge regions, a new discriminant function is constructed, and the segmentation threshold is selected by maximizing the new discriminant function. The validity of the proposed method is verified by experiment on various images with unimodal, bimodal, multimodal, or comb-like histogram patterns. The experiments on the noisy images show the robust anti-noise performance of the proposed method.Transition-based thresholding methods have the potential of segmenting the images with different histogram patterns. However, the existing transition-based thresholding methods have two obvious shortcomings including sensitive to noise and feature parameter. To overcome these disadvantages, we propose a novel thresholding method based on stable region set. The basic idea lies in that the different reasonable transition regions can form a transition region set. The elements of the set have higher stability in the average level. We utilize this kind of stability to distinguish the reasonable transition regions from the unreasonable ones. The average gray level of the middle element of the reasonable transition region set is treated as the final segmentation threshold. The experiments on many synthetic images and real ones show that the proposed method produces more accurate transition region and obtains better segmentation results.The third thresholding method proposed in this thesis is based on the pattern matching. The proposed method regards the part edge pixels extracted from the original gray level images as the reference model, and it determines the reasonable segmentation thresholds by searching the thresholding result which has the optimal match with the reference model. The proposed method adopts a new arctangent Hausdorff distance (AHD) for the pattern matching. The validity of the proposed method is verified by experiment on many images with various histogram patterns.Two main problems with watershed segmentation on the structure texture image are its inherent over-segmentation and under-segmentation. For overcoming these problems, we propose a watershed segmentation method based on combining control strategy (CCS) for extracting the objects of interesting. In view of its representativeness and actual engineering requirement, the skin texture image is taken as an important example in the proposed method. The proposed method achieves the segmentation by the cooperation of data-driven control and model-driven control. The experimental results on 60 skin texture images and other structure texture images illustrate the validity of the proposed method.