基于cuda的遥感图像运动目标检测-电子与通信工程专业论文.docxVIP

I摘要 I 摘要 摘要 基于遥感图像的运动目标检测是遥感领域的重要研究方向，检测实时性一直 是研究的热点话题。遥感图像数据具有数据存储量大、分辨率高、信息量多等特 点，而运动目标的检测算法（如数学形态学算法）的运算量大、复杂度高，基于 现有 CPU 串行计算方法的运行时间长，运行效率低，一般达不到实时性要求，亟 需寻求一种新的方法实现处理。 本文所采用的运动目标检测算法其核心算法（膨胀、腐蚀、开闭、灰度量化、 线性插值等）具有每个像素点的运算通常只与局部像素点相关、不同数据的具体 处理过程基本相似的特点，非常适合并行处理。而近些年发展起来的 GPU 通用计 算就是一种高度强调并行化的计算理念，能够很好的应对重复大数据量运算处理 的需求。优化得当，计算加速比能够是 CPU 的百倍乃至更高。CUDA 技术是 NVIDIA 公司推出的一种 GPU 通用计算编程模型。本文在分析最新图形处理单元（GPU） 的并行架构和统一计算设备架构(CUDA) 灵活的可编程性的基础上，将遥感图像 运动目标的检测算法在 CUDA 平台上进行了并行化处理，实现了运动目标检测的 高计算精度和高度实时性。 关键词：遥感图像 运动目标检测 数学形态学 CUDA 并行 IIIAb III Abstract Abstract Moving target detection in remote sensing image is an important research direction in remote sensing field, and its real- time detection has always been a hot topic. Because remote sensing image data has large data storage capacity, high resolution and huge scale information and the moving target detection algorithm such as mathematical morphology algorithm holds large computation and high complexity, the current CPU serial calculation method leds to long operation time and low operation efficiency thus can not reach the real-time requirements. It is urgent to seek a new processing method. In this paper’s main algorithm of moving target detection algorithm (expansion, corrosion, open and close, gray quantization, linear interpolation, etc.) the operation of each pixel is usually only related with local pixels and the basic process of different pixel data are similar, which makes it very suitable for parallel processing. GPU general calculation is a kind of concept highly stressed deserialize calculation, which could be used to deal with repeated large data quantity computing. If optimized proper, the computing speed could be one hundred times and even higher than that of CPU. By the discussion of the parallel architecture Graphics processing unit (GPU) and unified computing equipment structure (CUDA) with agile programmability, the moving target detection in remote sensing image on CUDA platform is realized achieving dat