Particle Swarm Optimization And It's Application In Image Processing

Swarm intelligence algorithm is a kind of stochastic algorithm based on the behavior simulation of biology swarm, such as ants and birds. The ant colony optimization algorithm, ACO and particle swarm optimization algorithm, PSO provided by M. Dorigo and J. Kennedy respectively are the two examples of swarm intelligence algorithm. Recently, swarm intelligence algorithm research on various engineer optimization problem has got more and more good results. The swarm intelligence optimization algorithm has become a hotspot in artificial intelligence, economic, sociology, biological sciences, computer science research area, etc. The study on algorithm principle and algorithm improvement can improve not only the optimization performance but also the possibility of swarm intelligence algorithm when it is used in the large scale combinatorial optimization problems.This paper focuses on the quantum-behaved particle swarm optimization algorithm(QPSO). By analyzing the principle of the algorithm, it gives out some improved versions to avoid the premature convergence problem. Meanwhile, in order to speed up the algorithm running process, some parallel QPSO algorithms are studied in detail. Experiments on benchmark functions and image process problems show that the parallel algorithm shows better performance in many optimization problems. In this paper, the improvement and parallelization research of the algorithm includes:(1) When the algorithm is applied to solve multimodal problems and complicated optimization problems with local optimum, the mbest in the algorithm will cause particles convergence to the local optimum. It will lower the Algorithm's performance in convergence speed and search ability. To overcome this problem, an improved algorithm named neighborhood topology QPSO(NQPSO) is proposed. By adjusting the neighborhood of particles dynamically, the improved algorithm maintains several attractors to avoid the premature convergence and enhance the particles search ability. The experiment shows that the improved algorithm has better performance than QPSO and SPSO in collective diversity and search ability, especially in solving high-dimensional optimization problems.(2)In order to avoid the premature convergence in the algorithm, an improved version is proposed. By introducing the Gauss disturbance on the mbest in the algorithm, the collective's activity and diversity are maintained when algorithm is running. Experiments on benchmark functions show the improvement on the algorithm with Gauss disturbance.(3) Another improved algorithms are studied. They are multi-phased QPSO, named MQPSO and diversity-maintained QPSO algorithm, named DMQPSO. In the first algorithm, the collective is divided into several sub-groups, the search procedure is divided into different phase, thus make the particles maintain the search ability when algorithm is running. The second algorithm keeps the particle's activity by controlling the collective's diversity to some degree. These two algorithms can avoid premature convergence and improve search performance effectively.(4) By studying the parallelization of the algorithm, a parallel algorithm with island modal is introduced in detail. In this algorithm, particle collective is divided into several sub-groups according to the computing nodes number in parallel system. Each sub-group of particles searches independently in different computing nodes. By exchanging search results between different sub-groups in period, the algorithm can keep the diversity of particle collective and improve the search performance. Meanwhile, a parallel computing modal based on dynamic neighborhood topology is proposed. A parallel version based on neighborhood modal is programmed by MPI, OpenMPI and MPI+OpenMP respectively. Experiments on benchmark functions show the good performance of this parallel algorithm.(5) The application of the algorithm in engineer optimization problems such as image registration and image segmentation is studied. Experiments also show the good performance of the algorithm.At the first section of this paper, we introduced the research background, research objects and some evolutionary algorithms in detail. In chapter 2, the principle of PSO and QPSO algorithm are described, and give out the realization method of these two algorithms. Chapter 3 focuses on the convergence speed and search ability. The principle of the algorithm based on neighborhood topology algorithm is studied and tested on benchmark functions. In chapter 4, the Gauss disturbance is used to improve it's diversity. In chapter 5, two improvements are applied to the algorithm, named a multi-phased QPSO and a diversity-maintained QPSO. The test results on benchmark function are given out in detail to show the performance improvement. In chapter 6, the parallelization of the algorithm is studied, including parallel computing model, parallel algorithm, etc. Chapter 7 describes the application of the algorithm and parallel version in digital image process. At the end of this paper, there are the conclusion of this paper and the future works.