Research On Image Thresholding Methods Based On Statistic And Spectral Graph

Image segmentation is a key step in image processing and analysis, and a classic difficulty in computer vision. There are some applications of image segmentation, where gray levels of object pixels are distinctive from those of background ones. In this case, image thresholding becomes a simple and effective image segmentation approach. During last few years, image thresholding has gotten wide attention from researchers at home and abroad, and has been widely applied to a lot of fields, such as target recognition and machine vision. The paper does comparatively deep research on image thresholding, and its main works and research results are as follows:(1) Classic statistical thresholding methods use variance sum of object and background classes as criteria for threshold selection. They only take variance sum into account, and fail to achieve satisfactory results when segmenting a kind of images, where variance discrepancy between the object and background classes is large. To solve the problem, a new statistical thresholding method combining variance sum and variance discrepancy is proposed in this paper. In addition, we present another statistical method for some images having similar statistical distributions on the object and background, and relate it with isoperimetric constant of a graph. This further shows the rationality of our method, and experimental results on a series of infrared images demonstrate its effectiveness.(2) Recently, image segmentation technique based on spectral graph is a new research hotspot. It regards an image as a weighted undirected graph, converts image segmentation problem into graph partitioning one, and implements image segmentation by minimizing certain cost function of graph partition. Among this kind of methods, isoperimetric cut is a newly developed one. However, the isoperimetric cut dose not belong to thresholding, and fail to adequately utilize gray level information of an image. This makes it unsuitable for gray level image segmentation. Here, we introduce the isoperimetric cut into image thresholding, and present a bilevel thresholding method for overcoming the above limitation. The proposed method uses isoperimetric ratio of the isoperimetric cut as criterion for threshold selection. Furthermore, characteristics of human visual perception are also utilized to reduce search range of thresholds, shorten segmentation time, and improve segmentation performance. The above bilevel thresholding method can only divide an image into two parts, and can not meet some practical segmentation tasks dividing an image into multiple parts. To solve the problem, we extend a bilevel method based on isoperimetric cut into multilevel thresholding. The extended method finds multiple thresholds by a fast and effective iterative scheme, simplifies computation of isoperimetric ratio, and introduces a way of automatically determining cluster number to adaptively choose reasonable threshold number. The new multilevel method can automatically determine threshold number, and its time complexity is independent of the threshold number. This makes our method avoid disadvantages of conventional multilevel thresholding ones, i.e., instability of segmentation performance and exponential growth of computational complexity with the increase of threshold number. Experimental results on a series of images show the effectiveness of our multilevel method.(3) Image thresholding based on transition region is a newly developed image segmentation technique. As compared with non-transition region (i.e., object and background regions), transition region of an image has more frequent and stronger gray level changes. On the basis of the characteristic of transition region, a transition region extraction and thresholding method based on gray level difference is proposed in this paper. The proposed method uses absolute difference between a pixel's gray level and the gray level average of its local neighborhood window as a descriptor for depicting transition region. The above gray level difference is very rough, and can not reflect the detailed difference between the pixel and other pixels in its neighborhood. Hence we present a modified gray level difference as a new transition region descriptor. The descriptor uses the sum of absolute gray level difference between the pixel and each pixel in its neighborhood to characterize transition region. Experimental results on a variety of images show that the modified gray level difference is effective on transition region description. In addition, conventional transition region descriptors do not take frequency and degree of gray level changes into account simultaneously, and fail to depict transition region comprehensively. To solve the problem, we present a new descriptor integrating local complexity and local variance. It uses local complexity to reflect frequency of gray level changes in local neighborhood window, meanwhile utilizes local variance to degree of the changes. Then local complexity and local variance are combined as a new transition region descriptor after being normalized respectively. Experimental results on a variety of images including infrared and text ones show that the new descriptor can depict transition region more accurately, as compared with conventional ones. And the corresponding method extracts transition region more accurately, obtains better thresholding results, and has stronger noise immunity.(4) Existing image thresholding methods based on transition region do not consider characteristics of human visual perception. An unsupervised transition region extraction and thresholding method is proposed to solve this problem. The proposed method first utilizes characteristics of human visual perception and statistical characteristics of an image to estimate gray level range of transition region for implementing image transformation in an unsupervised way, then uses local variance as descriptor to extract transition region, and finally obtains thresholding result. Experimental results on a variety of images including industrial nondestructive testing ones show that image transformation preserves gray level changes of transition region, meanwhile weakens gray level changes of non-transition region. This simplifies the original image, which should be helpful for subsequent transition region extraction. The new method obviously improves accuracy of transition region extraction, and obtains better sgemnetation results. In addition, we introduce the above image transformation into conventional thresholding, and present three unsupervised range-constrained thresholding methods. As compared with conventional approaches, range-constrained methods implement thresholding on the transformed image instead of the original one. This not only coincides with human visual perception, but also reduces search range of thresholds and saves computational time. The transformed image is simper than the original one, which should be helpful for subsequent image thresholding. Experimental results on a variety of images including nondestructive testing ones show that range-constrained methods have better segmentation quality, and segmentation speed is comparative with their counterparts.