Research On Parallel Algorithms Of Key Technologies Of Vehicle License Plate Positioning Based On GPU

License plate positioning is directly influence the correct probability of license plate recognition system and affect on the recognition speed. The most important thing to improve the efficiency and real-time ability of the vehicle license recognition depend on whether the license plate can be positioned quickly and accurately or not.On the other hand, the rapid development of semiconductor industry enhanced the performance of the computer system processor day by day. GPU, because of its more transistors to be used for internal data processing rather than process control, has more processors and more parallel process capacity overall comparing with the existing multi-core CPU. The modern GPU is regarded as strong large-scale parallel computing device. Therefore, general purpose GPU computing was proposed. It released powerful computing ability in many non-image areas. The programmable GPU give us a new idea to solve some common tasks efficiently. It can be seen as a high-performance device, fit to large-scale super compute based on desktop platform, and complete real-time analysis and mass data processing.CUDA( Compute Unified Device Architecture) platform is a new hardware and software architecture for implementation of general purpose GPU computing by NVIDIA. Programmers can write C-style code to start a lot of GPU threads in parallel based on it. So it provides a highly efficient and friendly development environment to solve many computation-intensive problems.This paper study on GPU-based parallel algorithm of the key technologies of license plate positioning and makes theoretical analysis and simulation. The main works are as follows:(1) Using GPU for image spatial filtering operation.Since the operation result of each pixel processed is independent, it can be done in parallel. This article parallelizes the process of binarization, edge detection and gray extended used in license plate positioning. Experiment results show that the efficiency improve extremely when the image was operated in parallel.(2) Using GPU for image geometric transform.Base on traditional transform algorithm, use position incremental offset method instead of a large number of multiplications of original algorithm and apply the rapid parallel computing ability of CUDA to image geometric transform, solve the problem of low efficiency of the image transform based on CPU.(3) Proposal a parallel algorithm for connected component labeling and implemented on GPU.It is difficult to label connected component in parallel for traditional algorithm because that it is dependent each other for each pixel operation result and it requires certain sequence. In this paper, base on the parallel architecture and hardware features of GPU, we proposal a complete parallel algorithm for binary image connected component labeling and get a better speedup.The results show that because of positioning license plate in parallel, it improved the procedure efficiency, and improved the real-time ability of the license plate recognition system finally.