【作者】 韩斌；

【导师】 王士同；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2009， 博士

【摘要】 计算机视觉检测是指利用计算机图像处理系统来模拟人类视觉进行检测对象的分类、识别和测量的技术,具有检测效率高、可检测指标多、可非接触测量、检测结果客观等优点。随着计算机视觉、模式识别、数字图像处理等技术的发展以及计算机硬件性能的提高,计算机视觉检测技术得到了越来越广泛的应用。不断涌现的应用需求,又给计算机视觉及相关理论的研究提出了新的研究课题。本文针对计算机视觉检测中粘连目标分离、动态目标特征描述、图像配准等问题进行了研究,并在实际计算机视觉检测系统中得到应用。本文的主要工作如下:(1)提出了基于模糊距离变换的粘连目标分离算法。以模糊距离变换与分水岭算法相结合,构造粘连目标分离算法,提高了目标边界模糊情况下的分离效果。(2)研究了离散情况下不变矩的比例不变性,纠正了相关文献的证明错误,推导出Hu矩的七个矩组在离散条件下φ2 ~φ7具有比例不变性、φ1不具有比例不变性的结论。提出了描述模糊边界目标的模糊边界不变矩,证明了其与清晰目标边界矩的等价性,对模糊目标提取边界不变矩特征时,可直接在模糊边缘带图像上进行,不需要首先提取其清晰边缘,方便了边界不变矩特征的提取,提高了边界不变矩特征提取的鲁棒性。(3)提出了基于投影矩的动态目标特征提取方法,将滑动窗内的若干帧动态目标图像叠加到一投影平面上,然后,采用投影图像的不变矩来描述目标的动态分布。与采用单帧图像特征相比,提高了分类的可区分度,同时具有较高的实时性。(4)利用克隆选择机制改进传统的量子遗传算法,增强了局部搜索的能力。将改进后的量子遗传算法应用到基于角点的待检测图像与模板图像的配准中,进行表面缺陷的检测。(5)介绍了上述方法在基于计算机视觉的蚕卵统计计数系统、船舶机舱火灾探测系统、印刷品印后缺陷检测系统中的应用。更多还原

【Abstract】 Computer vision detection is a technology of simulating human vision to classify, identify and detect objects detected by using computer image processing systems. It has advantages of high performance, more parameters, non-attachment, result-objectiveness and etc in detection. With the development of technologies in computer vision, pattern recognition, digital image processing and the promotion in computer hard wares, Computer vision detection techniques are getting more and more widely used in many areas. As application requirements appear continuously,they arise many new study topics for researching in computer vision detection and related theory. This dissertation studies problems about overlapping object separating, dynamic object characteristic description and image matching and their applications in real computer vision detection systems. The chief works are as follows:(1) It proposes an overlapping object separating algorithm based on fuzzy distance transform. Combining fuzzy distance transform and watershed algorithm, it constructs the overlapping object separating algorithm and promotes the separating effect in the case of object edge being fuzzy.(2) It studies proportion invariance of moment invariance in discrete condition,corrects the proof error in related literature, infers the result that in Hu’s seven moments,φ2 ~φ7 still keep scale invariance, butφ1 does not in discrete situation. And then, it suggests blurred boundary moment invariance to describe blurred objects and proves its equivalence to that of clear boundary object moment. Using this result, we are not necessary to precisely extract its crisp edge at first when extracting boundary invariable moment from blurred images. This makes it convenient to extract boundary invariable moment features and promotes the robustness of boundary invariable moment feature extracting.(3) It puts forward dynamic object feature extracting method based on project moment. This method firstly overlays the dynamic object images in several frames in the sliding windows onto a projection plane, and then it uses the moment invariance of the projection image to describe the dynamic distribution of the object. Compared with features in using single frame, it increases the discrimination degree of classification and is of higher practice.(4) It improves quantum genetic algorithm using clone selection mechanism and enhances local searching ability. Applying improved quantum genetic algorithm in the matching of a corner based image to be detected and a template image, we perform a detection with face fault.(5) It introduces applications of the above methods in a silkworm Ovum counting system, a ship cabin fire detection system and a printing matter fault detection system.更多还原