Study On Some Issues Of Image Thresholding And Object Segmentation Method

Image segmentation is a basic problem of computer vision, and it is the premise to achieve image understanding and recognition. Since it is simple, efficient and easy to implement, the thresholding approach has become one of the most important techniques of image segmentation, and it has been successfully applied in many fields, such as automatic target recognition, medical image analysis and industrial image processing. However, due to the complexity of the image, how to select the optimal threshold is still an unanswered question in image thresholding. Although many effective methods have been presented by many researchers in recent decades, the fundamental problem has not been solved yet. In addition, the problems of real-time and robustness in image processing tasks also are the trouble for thresholding method in practical application. So the presentation of thresholding methods that are efficient and suitable for image segmentation is still a challenging task at present. Firstly the state of art in image thresholding is reviewed and analized. Then, some studies and discussion on the theories and applications have been carried out in this paper, and some new ideas, methods and approaches for image thresholding have been introduced. In the object segmentation, the extraction of human object from infrared image was studied and two methods that combined with fuzzy sets theory, generalized entropy principle and chaos swarm intelligence optimization algorithms are presented. The main contributions of the thesis are as follows.(1) Based on the analysis of variance (ANOVA) principle and combining the information that provided by two-dimensional gray-level histogram of image and its local average image, two two-dimensional minimum class-variance thresholding methods are proposed. In order to reduce the computation time in searching the optimal threshold in two-dimensional histogram, the recursion and differential evolution algorithms are used in the new methods. The experimental results show that the proposed methods overcome the thresholding bias of the two-dimensional maximum between-class variance methods on the condition that the variances of the image foreground and background are not equal and improve the against noise performance of the one dimensional variance based methods.(2) Two thresholding methods are proposed based on the image gray-level com-occurrence matrix (GLCM) and the two-dimensional gray-level histogram of image and its local average image. One method uses image two-dimensional local cross entropy defined in GLCM and obtains the optimal threshold through minimizing the defined two-dimensional local cross entropy. The other method uses the two-dimensional gray-level histogram of image and its local average image. The two-dimensional gray-level histogram was projected to one-dimensional space firstly, and then the image is segmented by the minimum cross entropy method. The two methods improved against noise performance of the traditional minimum cross entropy method, and the last method obtains the similar computing efficiency compared with the one-dimensional method.(3) Image is a typical physical system with nonextensive property. However, the nonextensive information of image is usually ignored for optimal threshold selection in traditional thresholding methods. In the thesis, the nonextensive cross entropy which was presented by Tsallis is applied to image thresholding. The original formula of the Tsallis cross entropy is modified to be appropriate for image segmentation firstly. Then, two methods are proposed based on the vector distance metrics and the ability to handle the physical system with nonextensive property of the Tsallis cross entropy. The method that uses nonextensive property of the Tsallis cross entropy is extended to a two-dimensional form. The experimental results on a large number of synthesis and real images show that the proposed methods have better adaptability for complicated images.(4) Since the imaging process of the natural scenery and the image signal processing in image processing system have the fuzzy behavior in nature, all defuzzization methods for image thresholding have the defect on images with fuzzy property. To handle the fuzzy problem, a method combining the fuzzy sets theory and maximum correlation principle is proposed. Sometimes the object can not be extracted when the image is segmented by bi-level thresholding, so multi-level thresholding is necessary in this case. Therefore, a multi-level fuzzy thresholding method based on Renyi entropy and fuzzy sets theory is proposed. In order to accelerate the convergence of the proposed methods, the differential evolution algorithm is used. The experimental results show the effectiveness of the proposed methods.(5) The foundation of the infrared human object detection and recognition is the human targets extraction from the infrared image, i.e. the segmentation of the region of interest (ROI). However, the segmentation method of ROI is just in experimental stage and needs further exploration due to the unique characteristics of infrared thermal imaging. Two infrared human object segmentation methods based on Havrda-Charvát entropy, nonextensive entropy and fuzzy sets theory are proposed in the thesis. For accelerate the convergence of the object segmentation algorithms, two chaos optimization algorithms are presented when the chaos search is embed in classical swarm intelligence optimization algorithms, such as particle swarm optimization algorithm and differential evolution algorithm. At the same time, the mathematical morphological operators are used to denoise, fill cavity on the threshold segmented image for further image processing. The experimental results on real infrared images demonstrate the effectiveness of the proposed methods.Some theories and applications about image thresholding and target segmentation have been explored in this dissertation. A few criterion functions and algorithms based on some theoretical bases mentioned above have been presented which are more suitable for selecting the optimal threshold for image segmentation. The application of thresholding method to target extraction from infrared human image has been discussed. The proposed methods extend and improve the study in this area.