【作者】 骞森；

【导师】 朱剑英；

【作者基本信息】 南京航空航天大学 ， 机械电子工程， 2007， 博士

【摘要】 移动机器人在生活、国防和科学探索中的应用越来越广泛,而视觉系统是其智能化的重要实现手段。在室外非结构化环境中,图像退化是普遍现象,因此本文致力于解决视觉系统中的图像退化问题。首先研究了退化图像的质量评价,然后系统分析了退化图像的恢复技术,进而引入图像的退化不变量及其分析和应用。论文的主要内容和创新点如下:1.回顾了机器人视觉的研究历程,指出图像退化问题的重要性和现有研究的局限性。2.分析了非结构化环境中的图像退化类型,拓展了图像退化的含义,除了传统的灰度退化,还包含几何退化,具体包括图像噪声、图像运动模糊、光照变化与阴影、摄像机畸变、空间变换等;针对图像的这些退化,拓展了图像质量评价的概念,考虑了图像的不同尺寸,然后提出了基于奇异值分解和感兴趣区域加权的图像质量评价方法,并以实验证实和分析了其与人类视觉系统评价的一致性。3.拓展了图像恢复的含义,除了传统的灰度退化的恢复,还包含几何退化的恢复,然后系统分析了各种图像退化的恢复算法,用实验的方法分析了各自的优缺点,指出每种恢复算法的精度都有限,且需要退化类型的先验知识,否则图像恢复会造成不期望的结果;因而提出了图像的退化不变量的概念,即对各种图像退化都保持不变的特征量,这样避开了图像恢复的问题,直接把退化图像作为视觉系统的处理对象。4.分析了图像退化不变量的构造,并对一些传统的几何不变量,分析其退化不变性;实验表明Harris-Affine兴趣点、Hessian-Affine兴趣点和SIFT不变量可作为图像退化不变量;鉴于运动模糊的普遍性,特别提出了图像运动模糊不变量的概念,并对实验得到的几个图像运动模糊不变量进行了比较,指出各自的适用场合。5.分析改进了SIFT不变量的实现过程,减少了SIFT的计算时间;提出了基于SIFT的双向匹配算法,提高了图像匹配结果的正确率。6.通过实验分析了图像退化对机器人视觉中的立体匹配和目标跟踪的影响,并应用图像退化不变量理论来处理;实验显示了论文提出的退化不变量的有效性和工程价值,也指出了图像退化不变量在实际应用中需要改进之处。本课题的研究得到国家自然科学基金项目“智能微机电系统视觉/力觉/位移混合检测与控制技术”(项目批准号:50275078)的支持。更多还原

【Abstract】 Mobile robot has been more and more widely applied in life, defense and scientific exploring, and the vision system is an important implement technology of its intelligence. In outdoor unstructured environment, image degradation as a popular phenomenon is focused on in vision system in the dissertation. Degraded image quality measure is studied firstly, and then image restoration technoloty is systematically analyzed, based which image degradation invariant is introduced, analyzed and applied. The main contents and innovations are as follow:1. The research history of robot vision is reviewed and the importance of image degradation and the limitation of research on it is pointed out.2. The type of image degradation in unstructured environment is analyzed, and the definition of image degradation is extended to geometric degradation besides traditional photometric degradation, including image noise, image motion blur, illumination change and shadow, and camera distortion and space transformation, according to which the definition of image quality measure is extended to different image size. Single value decomposition and weighted region of interest based image quality measure method is proposed, whose consistency with human vision system is then proved and analyzed by experiments.3. The definition of image restoration is extended to restoration of geometric degradation besides traditional restoration of photometric degradation. Then restoration algorithm of different degradation is systematically analyzed, and both the advantages and disadvantages are analyzed and evaluated by experiments, concluding that the accuracy of each restoration algorithm is finite and prior knowledge of degradation type is needed, or some unexpected results will be produced. Consequently the concept of image degradation invariant is proposed, which is the feature invariant to all kinds of image degration. So the problems in image restoration are avoided, and the degraded images are direct processing objects in vision system.4. The construction of image degradation invariant is analyzed. The photometric degradation invariance of some classic geometric invariants is analyzed, and the experiments show that interest point of Harris-Affine and Hessian-Affine and SIFT invariant can be viewed as image degradation invariants. Since image motion blur occurs very often, image motion blur invariant is specially proposed, and some motion blur invariants obtained by experiments are compared, and the applying condition of each of them is proposed.5. The implement of SIFT invariant is analyzed and optimized, which decreased the computation time. SIFT based bidirectional matching algorithm is proposed, which improved the accuracy of image matching results.6. The effects of image degradation on stereo matching and object tracking are analyzed by expriments, and image degradation invariant theory is applied to solve the problems. The expriments demonstrate the effectivity and engineering value of image degradation invariant proposed in the dissertation. What image degradation invariant applied to engineering application should be improved is also proposed.The research is supported by project of National Natural Science Foundation“vision/ mechanics/ displacement combined detection and control technology for the intelligent micro mechatronic system”(Project No. 50275078).更多还原