Parallel Implementation Of The Key Algorithms Of Computer Vision

The development of computer vision has brought imperative demand on the performance for real-time applications. The traditional serial applications on a single core platform can not meet the demands. The emergence of multi-core and many-core platforms brings a new breakthrough for settlement of this issue. The parallel implementation of the key algorithm of computer vision on many-core platform has become the mainstream of the research. SMT-PAAG (Simultaneous Multithreading-Polymorphic Array Architecture for Graphics and image processing) is an array processor with simultaneous multithreading which has independent intellectual property right. SMT-PAAG has more highly efficient and flexible means of communication. It supports multiple parallel computing models, such as data parallelism, task parallelism and pipeline processing. SMT-PAAG provides a reliable platform for the parallelization of computer vision algorithms. Parallelization of the computer vision algorithms on SMT-PAAG is of great importance to the design and improvement of special processor for computer vision.This thesis discusses the parallel implementation of the key algorithm of computer vision in detail based on SMT-PAAG as well as the foundation of SD-VBS (The San Diego Vision Benchmark Suite). At first it introduces the developments and its applications of computer vision, the present situation of the parallelization of computer vision algorithm, and briefly describes the architecture of SMT-PAAG, gives details of the processing unit, communication mechanisms, execution modes and the simulator of SMT-PAAG Then it gives an introduction of SD-VBS and makes an intensive study of some key algorithms, including integral image, LBP (Local Binary Pattern) feature extraction and Harris corner detection and matching algorithm. Next it further elaborates parallel programming on SMT-PAAG, parallelization of the key algorithms is proposed and implemented by combining with the platform features of SMT-PAAG, especially focuses on the integral image based on pipeline and Harris corner detection algorithm using task parallelism. Finally, the results of the parallel algorithm are validated on the SMT-PAAG simulator; the experimental results are analyzed based on speedup and efficiency. The data shows that parallelization of the computer vision algorithm on SMT-PAAG has gained significant performance improvement.