Multi-objective Optimization And SAR Image Segmentation Via Artificial Immune System

The optimal problem is a fundamental issue in current engineering practice andscientific research. Thereinto, if there is only one objective function in this kind ofproblem, it is the single-objective optimization problems (SOPs), and when the numberof objective function is larger than one, it is the essential form of multi-objectiveoptimization problems (MOPs). This thesis is to focus on the buildings of the advancedmodel and theory of artificial immune system based multi-objective optimization, whichis the leading and promising field in the area of MOPs. The intensive study has beenimplemented in the following challenging subjects, including many-objectiveoptimization problems, new dominance scheme, adaptive clone strategy, and efficienttechniques in diversity maintaining. Finally, the advanced techniques and proposedimmune multi-objective optimization algorithms are successfully applied into SyntheticAperture Radar (SAR) image segmentation. The main contributions of the thesis can besummarized as follows:(1) The difficulty of current multi-objective optimization community lies in thelarge number of objectives. Due to lack enough selection pressure toward the Paretofront, classical algorithms are greatly restrained. To this end, an immune memory cloneselection algorithm is proposed to solve the problem of multi-objective optimizationwith a large number of objectives. The nondominated antibodies are proportionallycloned by their preference ranks, which are defined by their preference information. It isbeneficial to increase selection pressure and speed up convergence to the truePareto-optimal front. Solutions used to approximate the Pareto front can be reducedgreatly by preference information. Besides, an immune memory population is built topreserve the nondominated antibodies and ε dominance is employed to maintain thediversity of the immune memory population. Finally, the proposed algorithm performedeffective in testing several multi-objective problems with2objectives and DTLZ2andDTLZ3as high as8objectives.(2) The study of new types of dominance mechanisms is a hot and key issue incurrent EMO community, and ε dominance is a representative one among them.However, their ability in diversity maintaining is sensitive to different shapes of Paretofronts. This paper proposes an improved ε dominance mechanism by Isomap, whichemploys Isomap to embed the original population to low dimensional manifold space. The intrinsic geometric structure of them could be discovered and ε dominance isadopted to select data in the embedding space. Compared with traditional ε dominance,the mechanism does not lose valid solutions and can maintain a set ofuniform-distributed solutions. Besides, extreme-solution-check operator is proposed toenhance the ability of holding extreme solutions of ε dominance. The detailedexperimental comparison with NSGAII, SPEA2, NNIA and εMOEA shows that the twostrategies in this study are beneficial to uniformity and spread maintaining.(3) The efficiency of MO algorithms is highly related with the adaptability in thesearching process. Most traditional algorithms are always to assign all the solutions incurrent population to different ranks, which will induce the waste of computationalresources. An adaptive multi-objective optimization algorithm by online discoverednondominated solutions is presented for MOPs in our thesis. Here, three search phasesare devised according to the number of nondominated solutions in current population. Ifcurrent population contains very few nondominated solutions, global searching processis required and all the solutions need to build the dominance relations among them;when the population consists of adequate nondominated solutions, dominated onescould be ignored and the isolated nondominated ones should be allocated morecomputational budget for local search. To exploit local information efficiently, a localincremental search algorithm is proposed and merged into the model. This proposedalgorithm maintains the adaptive mechanism between the optimization processes by theonline discovered nondominated solutions, which has enhanced adaptability androbustness of the searching process.(4) How to devise a steady and high-powered MO algorithm? Here, we proposedan immune MO algorithm based on adaptive ranks clone and dynamic deleting scheme,in consideration of the searching process of current immune MO algorithm beingsensitive to the number of nondominated solutions and poor performance of diversitymaintaining. The adaptive selection scheme and adaptive ranks clone scheme by theonline discovered solutions in different ranks can enhance the robustness of theproposed algorithm, which is less possible to be trapped into local optimal Pareto front.Furthermore, it has been widely approved that one-off deletion could not obtainexcellent diversity in the final population; therefore, a m-nearest neighbor list (where mis the number of objectives) is established and maintained to eliminate the solutions inthe archive population. Finally, the proposed algorithm achieved satisfied results interms of convergence, diversity metrics, and computational time. (5) Considering the bad performance of SO with single clustering validity index indiscovering the complicated relations among image pixels, an artificial immunemulti-objective optimization framework is formulated for the first time and applied toSAR image segmentation. Moreover, a fused feature set for texture representation isconstructed and researched, which utilizes both the Gabor filter’s ability to preciselyextract texture features in low-and mid-frequency components and the gray levelco-occurrence probability’s (GLCP) ability to measure information in high-frequency.MO algorithms can discover the complicated relations among image pixels and have thegood partitioning performance in clustering the classification problems with differentgeometrical structures. Besides, the efficient and robust MO algorithm in AIS proposedin Chapter5of the thesis is used as the framework of the SAR image fine segmentation.Finally, satisfactory segmentation results of several texture images and SAR imagesfrom ERS-2satellite are obtained by the proposed method. The experimental resultsshow its great potential application in SAR image segmentation.(6) In view of weak performance of existing SAR image segmentation bysingle-objective evolutionary algorithms and very low efficiency of currentmulti-objective segmentation algorithms. A high efficient multi-objective automaticSAR segmentation algorithm in AIS is proposed in this chapter. It is essentially discretetwo-objective optimization problems for current SAR image segmentation. With this inmind, the multi-objective SAR image segmentation employ adaptive rank baseduniform clone and dynamic crowding distance deletion. The later scheme has beenviewed as poor performance in diversity maintaining for many-objective optimization,however it can obtain satisfied results for two-objective optimization and takes on lesscomputational complexity. Besides, SAR image filtering method is employed in theimage preprocessing stage in order to preserve the fine structure, details and texture forsubsequent objective of interest recognition and better image understudying. Theproposed algorithm obtained the obvious better partitioning results in segmenting twoartificial SAR images with many number of categories and real SAR images fromTerraSAR satellite.