| With the development of computer image recognition technology, modern societypresents higher requirements on the image recognition and analysis, only relying on theCPU computing era can hardly meet the increasing needs. Thus, CPU moves towardsmulti-core technology. However, it is still unable to meet people’s needs for a large amountof data calculations and increases the difficulty of programming. At this time, a parallelprocessing architecture hardware-Computer Graphics Processor (Graphic Processor Unit,GPU) has commanded increasing attention of the developers. Since the floating-pointcalculations speed of GPU is much faster than of the same period of CPU. Besides, internaldata bandwidth of GPU is also much wider than the bandwidth of CPU memory subsystem.These two points above makes the GPU more suitable for high-density of scientificcomputing than CPU. For a large amount of data calculation like marine steelplate three-dimensional measurement, GPU will have a great advantage. Developersincreasingly prefer to use GPU-based computer processing, especially in the process ofmeasuring the three-dimensional plate. Because of the large amount of data, developers useco-processor which is composed of GPU and CPU to process data instead of thetime-consuming CPU. Among them, CPU is responsible for processing the serial code andthe parallel code is handled in GPU, which can greatly improve the speed of operation andsave time resources.This paper introduces the powerful computing ability and the advantages of parallelprocessing of GPU and focuses on the current mainstream development platform CUDA,including the hardware architecture, software environment, programming model and thecurrent main application areas. Next, using GPU accelerates image preprocessingalgorithms in the time domain space and frequency domain space, which comprises imagesharpening algorithm, image smoothing algorithm and fast convolution algorithm etc. Forthe characteristics of the target steel plate texture feature is not obvious, some point-arraystructured light is projected on the steel plate via controlled projector by the IPC(IndustrialPersonal Computer) and then the paper selects the Harris corner detection algorithm andSURF local feature extraction algorithm to complete the extraction of the feature pointsbased on CUDA. Comparing the same calculation and complexity of the algorithms on CPUwith GPU, we can get the speedup ratio and verify the accelerated effects of GPU. Finally,this paper summarizes the CUDA programming optimization strategies. And for thecharacteristics of the small ship steel plate, the paper makes several optimization methods ofCUDA program. |
