GPU-Based Particle Image Fitting And Classification Algorithm's Research And Application

With the rapid development of modern society, the growth of information is rapidly. Every day weather forecasting, computational biology, business risk calculation, science research and other fields produce vast amounts of information. As device scaling and cooling technology's restriction, X86 architecture-based processors are difficult to have a big performance leap. To solve this problem, a lot of technologies are generated. Modern high-performance computers show that using the GPU to speed up computation has become a trend.This subject comes from the Institute of Biophysics, Chinese Academy of certain sections. In the study we observed vesicles within the virus particles to analyze the characteristics of the virus. Observed under the super-microscope in the ultra virus particles within the vesicle fluorescence images can be fitted by Gaussian function. Levnberg-Marquardt curve fitting algorithm is a stable and fast fitting method. When fitting a large number of vesicles under the images taken, this will be a difficult challenge.Considering he above issues with mass data and the latest technology of today's high-performance computing, this paper under heterogeneous structure of the CPU+GPU do the following contributions and work: 1. Design and implement CPU+GPU architecture based parallel GPU-GauseFitting program which fit for the full Gaussian function, non-angle Gaussian function, and circular Gaussian function. Experiments showes that when fitting image frames's number more than 1000, GPU program will get 40 times speedup compare to CPU-GauseFitting program.2. In response to the shooting situation to the two vesicles, we extend GPU-GauseFitting program. The expanded program GPU-GauseFitting is more universal. Experiments showes that when fitting image frames's number more than 1000, GPU program will get 60 times speedup compare to CPU program.3. Design and implement CPU+GPU architecture based Naive Bayesian classifier. When exam with above fitting data which's number is more than 1000, the result shows that GPU-NBC earns about 20 times speedup compare to CPU-NBC. When test from 10 datasets from UCI, the speedup will between 2 and 50.This paper's experiments show that CPU+GPU heterogeneous architecture based programs obtain good performance compare to ordinary programs. GPU can be used not only to accelerate high performance computer in the lab also apply to small and medium sized high performance computing.With the development of hardware and software technology, GPU will have a better universal.