【作者】 董鸿燕；

【导师】 沈振康；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2008， 博士

【摘要】 边缘包含了图像大部分信息,边缘检测是图像处理和分析的关键步骤,对后续高层次的特征描述、匹配和识别等有着重大的影响。本论文结合具体的项目“基于可见光卫星参考图与红外实时图匹配的图像制导技术”,对图像边缘检测技术及其应用进行了深入的研究。主要内容有:1、论述了边缘检测的基础理论,归纳并评述了边缘检测的几种主要方法和几种边缘检测算法评价标准,分析了边缘检测中存在的难点,指出当前边缘检测研究中存在的几个问题,并对边缘检测的发展趋势作出了展望。2、针对实际图像中边缘经常呈现的漫射边缘的情况,对漫射边缘的检测进行了研究。通过分析漫射边缘的特点和漫射边缘检测的难点,采用先去噪,再通过减小边缘模糊来增强边缘,最后用传统边缘检测方法进行检测的步骤来解决漫射边缘的检测问题。由此,本文提出了基于多级中值滤波和模糊加权均值滤波的模糊滤波算法来去除混合噪声。该方法用模糊推理对多级中值滤波和模糊加权均值滤波的结果进行合成,能在滤除混合噪声的同时,更好地保留图像的边缘细节。在增强边缘方面,从减小边缘模糊性的角度分析比较了几种具有代表意义的图像增强方法,针对形态边缘减宽增强算法的不足之处,提出了一种斜坡边缘的模糊减宽增强算法,在减小边缘宽度增强边缘的同时抑制噪声。3、多尺度边缘检测是边缘检测研究中的一个热点问题,也是其发展趋势之一。由于去除噪声和检测边缘密切相关,因此分别研究了多尺度边缘平滑去噪和多尺度边缘检测这两方面内容。小波优良的时频特性使得小波域的图像去噪方法得到了比较好的效果,然而,近几年提出的Contourlet变换具有比小波变换更强的边缘细节描述能力,由此本文提出了一种基于Contourlet变换的自适应阈值去噪方法,它能比传统的小波域去噪算法更有效地在去除噪声的同时保留图像的细节和纹理。在多尺度边缘检测部分,通过分析一般的拉普拉斯金字塔不适合进行捕获边缘奇异点,进而对拉普拉斯金字塔进行了改进,并在此基础上提出了一种基于改进拉普拉斯金字塔(LP)塔形分解的边缘提取算法,该算法能有效地检测边缘。4、为了提高边缘定位的精度,本文对亚像素边缘定位技术进行了研究。首先论述了亚像素边缘定位的由来、原理和前提条件,总结了现有的亚像素边缘检测方法。针对亚像素定位的一维和二维阶跃边缘模型,提出了一种基于Legendre矩的亚像素边缘定位方法。本文分别针对一维和二维边缘推导得出基于Legendre矩的亚像素边缘定位参数表达式。又由于三级灰度边缘模型能比二级灰度边缘模型更好地描述实际图像的边缘,因此又分析并推导了由二级灰度边缘模型产生的原理误差。然后从理论上分别分析了高斯白噪声对一维和二维亚像素边缘定位参数的影响。最后通过用仿真图像和项目所用的实际图像进行实验,说明该方法能有效地实现亚像素级的边缘定位任务。更多还原

【Abstract】 Because edge contains a majority of information of image, edge detection is a critical step for image procession and analysis, which has significant influence on the characteristic description, matching and recognition after it. Incorporating the specific project of "Guidance technology based on the matching of optical reference image and infrared real-time image", this thesis concentrates on image edge detection technique research and its application. The main contents include:1. The fundamental theory of edge detection is reviewed, then, the main edge detection methods and some algorithmic evaluation standards are summarized. Based on analyzing the difficulty of edge detection, several problems in current edge detection research are indicated and the edge detection technique’s direction is prospected.2. Considering that actual image edges are always diffuse edges, diffuse edge detection is studied. Through analyzing the characteristics of diffuse edges and the difficulty of diffuse edge detection, and then reviewing former diffuse edge detection methods, we smooth out the noise at first, then weaken edges’ fuzziness to enhance the edges, so the edges can be similar to step edges, and finally adopt the traditional edge detection method to solve the problems of diffuse edge detection. Therefore, a new image denoising method aimed at removing mixed noise while preserving edges is proposed in the paper based on multilevel median filtering and fuzzy weighted averaging filtering. This algorithm uses fuzzy inference to synthesize the results of multilevel median filter and fuzzy weighted averaging filter, which can better preserve the image edges while smoothing out the mixed noise. As to the aspect of edge enhancement, several typical image enhancement methods are analyzed and compared from the view of reducing the edge fuzziness. Then the shortcoming of morphological edge enhancement algorithm is discussed. Based on it, an adaptive fuzzy enhancement algorithm by edge width reduction is presented. This method can enhance the ramp edges by reducing edge width while smoothing noise out.3. Multiscale edge detection is not only a hotspot of edge detection research but also one of its developmental directions. Because edge denoising has a close relation with edge detection, multiscale adaptive filtering to remove noise and muliscale edge detection is respectively researched. Image denoising based on Wavelet transform has made a good effect for its good time-frequency character, however, the proposed Contourlet transform in recent years, has stronger capability than wavelet transform when describing the edge details. Therefore, an adaptive denoising algorithm based on the Contourlet transform is proposed. This algorithm can do better than traditional wavelet shrinking denoising algorithms in smoothing noise out and preserving edges and details. As to the aspect of multiscale edge detection, by analyzing that the common Laplacian Pyramid (LP) decomposition is not appropriate to capture the edge point singularities, an improved Laplacian Pyramid (LP) decomposition is used, and a multiscale edge detection algorithm based on Laplacian Pyramid is proposed. This algorithm can detect the image edges reliably and effectively.4. Subpixei edge location techniques are researched to improve edge location precision. The origin, principle and prerequisite of subpixei edge location are discussed and the current subpixei edge location methods are summarized. As to two step-edge models of one-dimensional and two-dimensional subpixei location, a subpixei edge location algorithm based on Legendre moments is proposed. The parameters’ mathematical expressions of subpixei location based on Legendre moments for one-dimensional and two-dimensional edge are respectively deduced. Because three-gray level edge model can better describe actual image edge than two-gray level edge model, the error because of using two-gray level edge model is analyzed and deduced. Then, effects of Gaussian noise on one-dimensional and two-dimensional subpixei edge location are respectively analyzed. Experiments on emulational images and actual images show that this method can efficiently implement subpixei edge location task.更多还原