Research Of Image Threshold Segmentation Based On Entropy

Image segmentation is a fundamental and important process in the field of computer vision and image engineering, whose core idea is to divide the image into several non overlapping and mutually independent regions according to some criteria.The segmentation quality will affect the follow-up work such as image understanding and image analysis. Due to the randomness of the image and diversity, at present there are a lot of different image segmentation methods. And the thresholding based segmentation technique has been widely recognized and applied in many corresponding research fields, since the algorithm is simple and computation complexity is small as well as the stable performance. It starts from the gray level histogram of the image, and then adopt a designed thresholding criterion to obtain the optimal threshold value of gray level for image segmentation. In this paper, from the point of view of image thresholding segmentation, the entropy based thresholding algorithms are discussed, including the following aspects.This paper introduces the basic theory of image segmentation, including the basic concepts of the entropy and image segmentation. The development of Shannon entropy based models and the corresponding thresholding algorithms are reviewed.The quality criteria of the segmentation algorithm are evaluated by parameter optimization and error statistics. By the analysis and comparison, we can find that the non extensive entropy is more applicable and flexible than the traditional Shannon entropy in image thresholding. This paper focuses on the possible application of nonextensive entropy in image segmentation.Firstly, Nonextensive parameter q is introduced by Tsallis entropy to illustrate the long-range correlation among elements of a set. It is widely used in image segmentation. In present work, the behavior of parameter q is analyzed by the mathematical principles as well as practical demonstrations. A reasonable range of q is then determined. For images that have no obvious correlations between object and background, two Tsallis entropies with different q values are adopted to describe thesets of object and background, respectively. A double q algorithm is proposed to detect the pixels’ gray-level correlations within these two sets and the correlation strength is described by tuning the q value. The effectiveness of proposed algorithm is demonstrated by the thresholding results of a series of infrared images. Furthermore,for an object with given gray-level correlation strength, the proposed algorithm can be used to trace and recognize the object automatically.Secondly, the maximum entropy segmentation algorithm based on gray level histogram distribution is easy to be disturbed by noise. In order to overcome this shortcoming and improve the accuracy and stability of the algorithm, we analyze the information contained in pixels including the gradient and the second order differential of the gray level. An "average gray level versus gradient sharpening" structure that consisting of two-dimensional histogram was constructed. Since the gray level distribution is independent of the gradient distribution, the two-dimensional histogram can be reduced into one dimension, and the image logic operation is adopted to repair the lost detail during the denoising process. The experimental results show that the maximum entropy segmentation algorithm can enhance the anti noise ability in the operation process. The real-time performance and accuracy of the algorithm is better than the traditional two dimensional segmentation algorithm. Moreover, the algorithm has good portability for different types of image segmentation.Within the framewrok of image segmentation, we illustrate the importance of nonextensive entropy in image thresholding. According to the properties of the images, the corresponding segmentation algorithms are proposed and the effectiveness is verified.