【作者】 陈慧；

【导师】 陈树中；

【作者基本信息】 华东师范大学 ， 系统分析与集成， 2004， 硕士

【摘要】 大多数图像处理任务都是计算密集型的问题，数据量和运算量大，给实时应用带来困难，而图像并行处理技术是提高图像处理速度的最有效技术。 在图像处理技术中，Hough变换在形状分析方面被公认为是一种强有力的的工具，即使在有噪声的情况下，它仍能给出比较好的结果。但它有两个主要的缺点，一是运算量大，运算时间长，不能满足某些实时性要求；二是耗费内存空间多。因此，它也需要并行技术。本文对它的并行化技术进行了深入的讨论。其主要工作如下： 1．图像并行处理技术的许多基本概念，都来自计算机并行处理的概念。因此我们首先对并行计算机的发展及其体系结构进行了回顾。另外，在并行算法方面，对其定义和分类作了介绍，并对并行算法的设计方法和性能评价进行了阐述。 2．接着本文介绍了Hough变换，并用三种不同的体系结构和方法来实现它的并行化。 3．首先在网孔连接的计算机上实现Hough变换。由于二维网孔计算机的结构比较简单，而且比较规范，使得它成为解决计算机图像处理问题首选的并行体系结构。本文针对传统Hough变换只能统计直线上黑色象素个数，不能记录线段的两端点坐标及长度的缺点，介绍了一种改进Hough变换算法，并在网孔处理机阵列中实现，还给出了能够识别并消除重复线段的并行算法，这些在实际问题中是很有用的。 4．尽管网孔类型的体系结构很适合图像处理算法，但网孔本身通信直径大，当处理长距离的数据传递操作时，速度比较慢，因此人们提出了由连续的尺寸递层减小的网孔组成的金字塔体系结构，它结合了网孔和树类型体系结构的特征。本文提出了一种有效的多分辨率Hough变换在金字塔机器中的流水线实现。多分辨率Hough变换使用一系列的多分辨率图像和累加器数组的特性，可有效地把它映射到金字塔不同层大小可变的处理器阵列中。这种方法充分地利用了处理器，并且流水线这种实现方式可适用于连续图像的有效分析。 5．并行计算机体系结构包括单指令流多数据流(SIMD)和多指令流多数据流(MIMD)两种结构。前面两种办法都是基于SIMD结构。本文介绍了一种新的包含了MIMD和SIMD两种结构的系统结构，我们称这种新的结构为混合系统。而且，我们介绍了一种适用于此混合系统的新的Hough变换定义方法，并在此混合系统中实现，取得了较好的效果。更多还原

【Abstract】 Most image processing tasks are very computationally intensive. The amount of data involved and computing power required are very large, which bring great difficulties in real-time applications, while image parallel processing technology is the most efficient technology to improve image processing speed.In the image processing technology, Hough transform is considered as a powerful tool in shape analysis which gives good results even in the presence of noise. But it has two major shortcomings. One is large computing power and long computing time, which can’t satisfy some real-time requirements, the other is excessive storage requirement. Therefore, it needs parallel technology. In this paper we give a further consideration about its parallel execution, and the work of the paper is included as following.Many basic concepts in the image parallel processing are from computer parallel processing, so we first look back on the development of parallel computer and its architectures. Otherwise, in the parallel algorithm, we introduce its definition and classification. And on the base of this, the design method of parallel algorithm and its performance evaluation are described.After presenting the definition of the Hough transform, we apply three different architectures and methods to implement parallel execution of the Hough transform.First we implement Hough transform on the mesh-connected computer (MCC). Because of its simplicity and regularity, the 2-dimension MCC is a preferred parallel architecture for solving the computer image processing problem. The conventional Hough transform can only count the number of feature points which are collinear, and can’t record the end points of a line segment and its length. Thereby, this paper present an improved Hough transform to overcome these two shortcomings. The algorithm is parallelized and implemented in MCC. What’s more, we give a parallel algorithm to identify and eliminate overlapping line segments. These advantages are very useful in the real applications.Although mesh type architecture is very suitable for the image processing, its communication diameter is very large. It tends to be slow when it comes to handling data transfer operations over long distances. So pyramid architectures, consisting of a stack of successively smaller sizedmeshes, have been suggested. Pyramid architectures combine the features of mesh and tree type architectures. This paper presents an etTicient pipelined implementation of the Multiresolution Hough transform (MHT) on a pyramid machine. The MHT uses a set of multiresolution images and accumulator arrays at the different layers of the pyramid, which can be efficiently mapped into the variable sized processor arrays at the different layers of the pyramid. Due to this, the method results in a high utilization of the processors. Besides, the pipelined implementation fashion is suitable for real-time analysis of a continuous sequence of images.Parallel computer architecture consists of two classifications: simple instruction stream-multiple data stream (SIMD) and multiple instruction stream-multiple data stream (MIMD). The above two methods are both based on the SIMD structure. In this paper, we introduce an architecture, referred as Hybrid system, combines both SIMD and MIMD system. Furthermore, we present a new parallel-based Hough transform for implementation on the Hybrid system and give a better result.更多还原