| Image segmentation is an important step in the analysis of image data. For itsexcellent characteristic of simplicity and effectiveness, image thresholding has alwaysbeen a hot topic in the application and research of image segmentation. It segments theimage into regions without intersection by closed and connected edge through simpleinput of gray scale distribution. It is widely used in the field of document imageprocessing, industrial nondestructive testing, medical image processing, infrared imageanalysis and other fields. However, though lots of thresholding methods were proposedover the last decades, the adaptability of thresholding algorithm was not resolvedfundamentally. In addition, with the increment of the information considering and theapplication of new technology in thresholding, the time complexity of thresholdingalgorithm increase greatly and became the bottleneck of practical application. Thisthesis investigate these problems based on the wide analysis of existing methods and thestudy of new technology in other field of image processing, its main work includes:1) The necessary conditions for Otsu method to get the optimal threshold isanalyzed theoretically, and then the essential cause of its bias is pointed out. Based onthis a thresholding method adapt to histograms with different distribution is put forward,the new method is suited for the segmentation of series images with similar histogramdistribution especially.Lots of literatures pointed out by experiments that when classifying two categorydata with big difference in their variances, the threshold obtained by Otsu method tendto deviate to the data with big variance. By theoretical analysis of the solving process ofOtsu algorithm, the necessary condition for it to decide the optimal threshold is obtained,and thus the essential cause of the deviation of Otsu algorithm and its relative methodsis pointed out. On this basis, an adaptive threshold method based on minimum with-inclass exponential variance is put forward and extended to two dimensions. And thus theadaptability of the method is improved by the exponential parameters which adjustadaptively according to the distribution of histogram. A fast calculation method basedon cumulative matrix is combined with self-adaptive particle swarm algorithm to realizequick selection of image threshold. Compared with Otsu method and its twoimprovements on both synthetic and real images, our method is demonstrated to bemore accurate and adaptive. 2) A kind of GSGECM(Gray level-Strength of Gradient Excitation Co-concurrenceMatrix) is constructed following the idea of Weber’s law. On this basis, a kind ofthresholding algorithm with good anti-noise performance is put forward. Then a fastcalculation algorithm of two-dimension entropy and an improved Shuffle frog leapingalgorithm are combined to realize threshold selection.The ignorance of the noise and the edges result in the low anti-noise performanceof traditional division of two-dimensional entropy. Based on weber’s law, a kind ofGSGECM considering the direction and strength of gradient is put forward, and it iscombined with new division of two-dimensional entropy to improve the anti-noiseperformance of thresholding algrithm. And then, a kind of two-dimensional weightedRenyi entropy thresholding algorithm is put forward based on GSGECM. A new parallelintegral matrix calculation method is put forward to reduce the time complexity oftwo-dimensional Renyi entropy to constant. An improved SHLA with fuzzy leapingstrategy based on polar coordinate is put forward to implement parameters optimization.Comparative experiments on both synthetic images and real images illustrate thatanti-noise performance and visual effect of the proposed method is better than thecompeting methods.3) Considering the fuzziness and the non-additivity of digital image information, amulti-level threshold algorithm based on fuzzy Armioto entropy is put forward, in thesame time, chaos disturbance is added into quantum genetic algorithm to realize fastmulti-level thresholding.Image information is fuzzy and non-extensive, the fuzzy entropy based onShannon entropy is only applicable to additive system. Considering Armioto has betterperformance in the description of non-extensive information in images, the histogram ofimage is translated into fuzzy domain on which the fuzzy Armioto entropy is defined.Considering the complexity of natural images, just thresholding images into object andbackground cannot meet the demand of practical application sometimes, a multi-levelthresholding method based on fuzzy Armioto entropy is put forward to divide imageinto background, middle region, and object, and then the fuzzy Arimioto entropy ofthem is defined respectively, as well as the pseudo additive formula of three subsystemsis derived, and the total entropy of image is calculated according to this formula.Chaotic disturbance is added into QGA (Quantum Generic Algorithm) to help it escapefrom local extremum, and this improved QGA is employed to search the optimumparameters combination of the fuzzy membership function, and thus the thresholding of image is realized. Compared with typical thresholding method on real images and thenanalyze their results by both visual inspection and objective measure, it is demonstratedthat the proposed method is generally better that competing method.4) To address the problem of fuzzy imaging and low SNR of infrared images, Athresholding algrithm of infrared human body based on two-dimensional fuzzy Tsallisentropy is proposed, and a fast algorithm is designed to reduce the time complexity oftwo-dimensional fuzzy entropy from O(L2) to O(L).Limited by infrared imaging technology, infrared sensor generally has lowerresolution, higher fuzziness, and lower SNR, therefore, when the existing1D fuzzyentropy method and2D crisp entropy method are applied in the field of infrared humanbody thresholding, good results cannot be always obtained. Both the fuzzy nature ofinfrared image and non-extensive information widely existed in images are consideredin this thesis to address the above problem as follows. First, a kind of fuzzy Tsallisentropy is defined based on fuzzy partition rules. Secondly, in order to make full use ofthe special information of pixels to deal with noise problem, using fuzzy relation theory,the2D histogram is transformed into fuzzy domain and thus corresponding fuzzy subsetis produced, based on which the fuzzy Tsallis entropy is extended to two dimension. Toovercome the huge calculation burden brought about by extending1D fuzzy entropy to2D, a fast algorithm is put forward to reduce the time complexity of2D fuzzy entropyfrom O(L2) to O(L). Finally, combining the fast algorithm with chaotic QGA, thesegmentation of typical infrared human images is realized. Experiment on both opendatabase and images captured by ourselves are carried out, and the results are analyzedby both visual inspection and quantitative analysis of absolute error, as well as thereal-time performance is analyzed by both time complexity and CPU runtime, theconclusion is that the proposed method can obtain better segmentation result and higherreal-time performance than competing method. |
PDF Full Text Request