A Study Of Parallel Evolutionary Algorithm And Its Application In Image Processing

Evolutionary algorithm is a class of global optimization search algorithm based on mechanism of biological evolution, such as natural selection and genetic mechanism. Evolutionary algorithm has been widely used in digital image processing because it is simple, strong robust and parallelizable. However, evolutionary algorithm has a high time complexity and a low running speed. With the continuous improvement of digital is imaging technology, the amount of data contained in a digital image dramatically improved, creating several new challenges for massive image processing by using evolutionary algorithm. Represented by GPU and MIC, many-core computing devices have a large number of computing units and huge computing power. Parallel algorithms based on many-core computing devices have a great advantage in processing speed, compared with traditional serial algorithms. In this paper we parallel two kinds of evolutionary algorithms to duel with digital image processing problems by Open CL. The main work is as follows:We apply immunodominance theory in image feature selection problem, and design a unidirectional mutation strategy. To solve the time consuming problem for the serial algorithm, this paper presents a parallel immunodominance clonal selection algorithm(PICSA) to solve image feature selection problem. Compared with traditional algorithms, PICSA is higher in searching performance, faster in convergence and less in operation time. We use master-slave parallel model of evolutionary algorithm and design the strategy of random number generation on Open CL device, reduce the communication between the host and the device significantly. For multi-GPU device, propose an island parallel model based on multi-GPU devices in order to combine the parallel algorithm and hardware architecture efficiently. Parallel several steps which are not suitable for parallel processing previously by parallelization operators, and further improve the parallelization of the algorithm. The experiment on several data sets show that, the dimensions of the feature subset are higher in classification accuracy rate, lower in dimension, and a considerable speedup is achieved on each data set. Finally, the parallel algorithm is implemented on multi-core CPU, GPU and MIC to study the portability of this algorithm.Using clustering idea to solve image change detection problem, including the noise of remote sensing image and cluster centers chosen. So, we combine Particle Swarm Optimization and local information and in order to solve the time consuming problem for the serial algorithm, a new parallel clustering algorithm PPSOCLI is proposed. Several experiments on remote sensing image show that change detection results are higher in classification accuracy rate and fewer in iterations. We study on acceleration effect of the proposed algorithm in different range of local information, number of iterations and size of image. And acceleration effect in different number of threads by Open MP parallel algorithm.