【作者】 谢永胜；

【导师】 余正生；

【作者基本信息】 杭州电子科技大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 数字图像拼接与配准技术作为图像处理的关键环节,正成为图像处理领域研究的热点。其应用领域相当的广泛,如人脸识别、视频监控图像分析、银行印章识别、医学诊断以及全景图生成等等方面有着非常普遍应用,在科研领域内得到了极大的重视和发展。图像拼接与配准技术经过多年的发展已经积累了大量的研究成果,特别是配准技术,作为拼接的基础,相关技术也达到了较高的水平。图像拼接的方法主要有基于区域和基于特征,基于特征的拼接方法,主要是找图像之间的对应特征点,可以是点、边、曲率等特征,其优点是计算量比较小,但不够精确;基于区域的方法直接对每个像素进行比较,只要时间足够长,则可以拼接的足够精确,由于是对整个图像的像素操作,因此计算量特别大。配准技术主要是分为基于灰度、基于特征以及基于决策树这几种情况。研究最多的是基于灰度的图像配准,这种方法直接利用图像的灰度信息进行配准,通过像素对其间某种相似性度量的全局最优化实现配准,这种配准方法配准的速度比较低,对灰度信息变化非常敏感,没有充分利用灰度统计特性,对每一点的灰度信息依赖较大;基于特征的图像配准方法的研究也逐渐增多,这种方法先提取图像的特征,然后利用图像的特征进行配准;基于决策树的图像配准需要先建立自动表决专家系统模型,然后根据一定的决策规则进行配准。对这种方法的研究,目前没有突破性进展。本文主要是讲述对于图像拼接以及配准技术方面的改进。对于图像拼接技术,如何比较准确快速地定位好衔接重叠的区域是体现其算法是否高效的一个重要指标。因为在图像拼接过程中,比较大的一部分时间主要是消耗在寻找相匹配的区域上面。针对基于区域拼接技术在定位重叠区域耗时长的特点,引入了利用多线程技术以及改进的SSDA算法,来提高定位的效率。通过与一些传统算法的比较,速度方面有很大的提高。本文在配准的改进方面,则根据图像迹算法配准率高,以及分形维配准不需要提出图像的边缘、轮廓、纹理等特征,对于图像边缘、轮廓、纹理信息不丰富的图像也可以进行配准的特点,提出了一种提高配准率的方法,就是在传统的分形维配准技术的基础上加入了图像迹的技术,另外还引入了噪声处理方法,在配准时间也有很大改进。更多还原

【Abstract】 The development of computer graphics, is producing more and more far-reaching impact. Digital image stitching and registration technology as a key link in image processing has become a hot research field of image processing. Digital image registration fairly wide range of applications, such as face recognition, video surveillance image analysis, bank seal identification, medical diagnosis and panorama picture formation etc., has a very popular application, so this area of scientific research has been greatly attention and development.Image stitching and registration technology have accumulated a large number of research results over the years, especially registration technology, as the basis for stitching, related technology have achieved a higher level. The methods of image stitching are mainly region-based and feature-based, Feature-based stitching method, is mainly to find the corresponding feature points between the images, which can be points, edges, curvature and other characteristics,it’s quantity of calculation is small, but is not pricise enough. Region-based approach is to compare each pixel directly, as long as time is long enough, you can splice a sufficient precision. Because it is the entire image pixel operations, so computation is particularly large. Registration techniques are mainly divided into basing on the grayscale, feature and decision tree. The most studied is based on gray-scale image registration, this method directly use the gray-scale image registration information, through the pixels on the similarity measure during a global optimization to achieve registration. This registration method is relatively low rate of registration of changes in gray level information is very sensitive to not make full use of gray-scale statistical properties of each point of larger gray-scale reliance on the information. Feature-based image registration methods of study gradually increased, this method first extract the image features, and then use the characteristics of the image registration. Image registration based on the decision-making need to set up an automated expert system model of the vote, and then the decision-making rules based on certain alignment. The study of this approach, there is no breakthrough.This paper is mainly about the technology improvements of image stitching and registration. For image stitching techniques, how to locate a good convergence of overlapping area more accurately and quickly is the embodiment of their algorithm as well as an important indicator of efficiency. Because in the image stitching process, the relatively large part of the time is mainly consumed in looking for the match area. For the time-consuming characteristics in positioning the region-based stitching overlap proposed the use of multi-threading technology, and improved SSDA algorithm to improve the targeting efficiency. Compared with some traditional algorithms, the speed has greatly improved. The improvements in the registration, according to the high registration rate of image trace, as well as the fractal dimension of the image registration do not need to make the edge, contour, texture and other features, for the image which edge, contour, texture information is not rich can also be carried out, this paper put forward a method to improve registration rates in the traditional registration technology on the basis of fractal dimension by adding the image trace technology, furthermore the introduction of noise processing methods, registration time also improved significantly.更多还原