【作者】 吕常魁；

【导师】 姜澄宇； 王宁生；

【作者基本信息】 南京航空航天大学 ， 机械制造及其自动化， 2005， 博士

【摘要】 AVI缺陷检测已经成为现代工业生产与质量控制的一个重要组成部分。我国AVI的发展起步较晚,目前国内企业所应用的AVI系统仍以国外产品为主。发展具有自主知识产权的、适合我国国情的AVI系统,有着重要的现实意义。 以软件构架为主、基于PC机的AVI系统以其经济、灵活的特点,已逐渐成为国内厂家为新设备配套和进行旧设备改造的认可方案。实时性、鲁棒性与准确性是对AVI系统的基本要求,基于PC的AVI系统的性能优劣很大程度上取决于所采用的计算模型。本文针对药品生产中片剂和胶囊制剂铝塑泡罩包装的缺陷检测,将跟踪机制引入到药品包装AVI中,提出了“运动检测—跟踪—缺陷检测”的检测模式,并围绕这一结构模式,对运动检测、多目标跟踪以及规则纹理的缺陷检测三个方面的关键算法技术进行了研究。 运动检测是实现跟踪的前提。文章基于只检测运动物体边缘轮廓的思想,构建了基于块区域变异系数最大似然比的三帧差分MLR运动检测模型,同时采用基于核密度估计的背景减除模型来滤除MLR模型检测结果中的非运动区域,建立了鲁棒的MLR-KDE自适应混合模型;针对MLR-KDE模型所检测到的运动区域,提出了能够对不连贯区域进行快速轮廓追踪的RW算法,为实现多目标跟踪奠定了基础。 将图论的方法与SVC聚类相结合,采用只对多维数据集的最小生成树主干进行递归连接运算的方法,对传统的SVC聚类算法进行了改进,以在保证聚类质量的前提下实现对小数据集的实时聚类;并将改进的聚类算法应用到对MLR-KDE模型所检测到的多目标运动区域的实时动态分割中,获得了较为满意的效果。根据运动区域的分割结果,结合RW轮廓追踪算法,采用基于特征的跟踪方案,提出了隔帧法实现多目标跟踪的思想,构建了相应的计算模型。 片剂和胶囊制剂通常具有规则的纹理表面。文章以规则纹理图像为研究对象,构建了纹理缺陷检测的统一模型。在对纹理图像进行小波多分辨率分析的基础上,采用自相关函数分别对多分辨率分析结果的LH和HL频带进行局部统计性描述,提出了基于自相关序列Fourier频谱分析的子带规整度的概念,给出了相应的表达模型,并进一步定义了多尺度纹理局部规整度的向量表达形式;更多还原

【Abstract】 AVI (Automated Visual Inspection) are now playing an important role in contemporary industrial manufacturing and quality control. In our country, development of AVI technology is still in its infancy. The AVI systems applied by domestic enterprises are still relying heavily upon import products. It is strategically important to develop the AVI systems with self-owned intellectual property rights to meet the needs of domestic market.PC-based, software-dominated AVI systems, with the advanced features of economy and flexibility, have been gradually accepted by domestic enterprises, for the revamp and complement of facilities. An AVI system should be robust, fast and reliable. Many applications demand inspection tasks operate in real time. The performance of a PC-based AVI system relies mostly on its computational model. In this thesis, for the defect detection of tablets and blisters in pharmaceutical packaging, we introduced the target-tracking mechanism into AVI and proposed a novel inspection model, during which a "motion detection-target tracking-defect detection" strategy is adopted for moving target inspection. Based on this model, we pursued an in-depth research on the algorithms of motion detection, multi-object tracking and texture defect detection, respectively.Motion detection is the basis of object tracking. Based on the idea only to extract the silhouette of a moving object, we created a new three-frame differencing model, say, the MLR model, based on the maximum likelihood ratio of CV (Coefficient of Variation) of successive blocked frame regions. Simultaneously, a kernel-based density estimation model is used as a filter to suppress false positives of the motion areas detected by MLR model. Integrating it with the MLR model, we constructed the hybrid MLR-KDE model, which is robust and self-adaptive. To fulfill the need of multi-object tracking algorithms in chapter three, we proposed a fast boundary tracing algorithm, namely the RW (Roller Wheel) method. This algorithm performs well even the processed regions are non-connected.更多还原