GPU-Based Parallel Artificial Fish Swarm Algorithm And Its Applications

Artificial Fish Swarm Algorithm (AFSA) is a swarm intelligence optimization method based on simulation of fish swarm behavior. In recent years, AFSA has been successfully applied to a range of problems, including power system, controller design, image segmentation and data clustering, among others. This algorithm has been proved to be a high-performance optimization method with robustness, fast convergence and capability of global optimization.However, when dealing with complex engineering problems, AFSA becomes quite time-consuming, limiting its scope of application and scalability. In order to improve the efficiency of the algorithm, parallelization is an important solution, which has not been used to this algorithm so far. In recent years, with the improvement of the programmable graphic processor, research on the GPU general computing gradually becomes active. To shorten the running time, this essay presents a GPU-based parallel implementation of AFSA (GPU-AFSA), which achieves a relatively rapid speed by converting the problem-solving-processes of artificial fishes into a batch of threads executing on GPU in parallel. Experimental data show that GPU-AFSA is able to achieve up to 30 times speedup, and has the same performance of optimization with the serial implementation of AFSA on CPU.Image segmentation, as an important step of target detection and identification, is a process in which targets of interest are retrieved from input image to output image. Image clustering is an important and widely used branch of image segmentation method. However, the algorithm of image clustering with broad applicability does not exist at present and most clustering algorithms used on image segmentation require large amount of computation. This essay presents an AFSA-based image clustering model integrating the high-performance optimization capability of AFSA. This model has high accuracy of segmentation, and is more adaptive than other image clustering methods with high dependency of characteristics of image data. Furthermore, two parallel implementations of this model on GPU are presented, both of which correspond to special requirements of image segmentation. Experimental results show that this model is an efficient and effective method with the capability of segmenting targets from different types of test images.