Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

Investigation on Applying Nonextensive Entropy to Image Segmentation

Author LinZuoZuo
Tutor OuCongJie
School Huaqiao University
Course Electronics and Communication Engineering
Keywords Image segmentation Thresholding method Tsallis entropy Nonextensive parameter Long-range correlation
CLC TP391.41
Type Master's thesis
Year 2013
Downloads 2
Quotes 0
Download Dissertation

Image segmentation is the most important and popular research topic in imageengineering. The basic operation of image segmentation is to segment the image intosome meaningful and non-overlapping regions. The results of segmentation candirectly affect the quality of subsequent image processing tasks. At present,researchers have proposed several segmentation methods for different types of image,of which thresholding method based on entropy is relatively simple and effective one.Tsallis Entropy is a generalized form of Boltzman-Gibbs entropy. It caneffectively describe the complex long-range interactions and the long-durationmemory within a nonextensive physical system. In the field of image processing,such long-range interactions (or long-duration memory) can be considered as thecorrelations among the pixel gray levels, and the strength of correlation can bedescribed by the nonextensive parameter q. Tsallis entropy-based thresholdingmethods are more universal and flexible than Shannon entropy-based ones whensegmentating some images with nonextensive characteristics.Most of the Tsallis entropy-based thresholding methods assumed that thelong-range correlations within the image are global. However, for some imageswhose objects can be arbitrarily changed, we can judge that there should not existlong-range correlations between the object and background by the prior knowledge,but there may be correlations within the object or the background. That is to say, thiskind of image has local long-range correlations rather than global long-rangecorrelations. Therefore, we improved the existing several Tsallis entropy-basedthresholding method. They are listed as follow.Firstly, we analyzed the deficiencies of the maximum Tsallis entropythresholding method. In order to overcome the deficiencies, we proposed a novelTsallis entropy thresholding method and analyzed the proper range of q value. Theexperimental results show the superiority of our method for some infrared images aswell as some nondestructive testing (NDT) ones, which have obvious local long-range correlation.Secondly, by constructing the two-dimensional gray histogram of the image, theabove-mentioned method can be generalized to two-dimensional one. In this case, notonly the pixels’ gray level information but also the neighborhood spatial relationshipsof pixels are taken into account. Therefore, this novel method can achieve goodsegmentation results even the signal-noise ratio (SNR) of the images are reduced.Finally, we analyzed the minimum Tsallis relative entropy thesholding method.By the same trick, we proposed a novel one to improve the performance. Theexperimental results show that this method can obtain better segmentation results thanthe minimum Tsallis relative entropy thesholding method when segmentating someimage in which the object and the background have no obvious correlations. Thisfurther show the validity of our assumption that there may exist local long-rangecorrelations among the gray levels of the pixels.

Related Dissertations
More Dissertations