【作者】 邓仕超；

【导师】 刘铁根；

【作者基本信息】 天津大学 ， 光学工程， 2010， 博士

【摘要】 随着光学显微镜和计算机技术的结合,光学显微镜正在向自动化和智能化方向发展。显微目标图像处理技术能极大地提高光学显微镜的性能和效用,对于光学显微镜的发展和应用有重要意义。本文根据光学显微镜的发展趋势和光学显微镜观测目标的特点,把观测目标分为固定形态目标和非固定形态目标两类,研究了与之相关的目标识别关键技术,这些技术包括:用于扩展显微镜高倍视场下景深的多聚焦图像融合算法,单像素封闭边缘提取算法,基于SVM的固定形态目标自动识别与测量系统,基于实时学习模式的非固定形态目标分类识别系统。并通过对比实验和性能实验检验了这些算法和系统的有效性和适用性。本文创新性工作概括如下:1、提出了基于自适应滤波器清晰度评估函数和根据最近邻权值判断融合条件的多聚焦图像融合算法。该算法可以有效地减轻噪声影响和增加显微镜景深,在多种倍率物镜下获得很好的融合效果,有利于后继图像处理。2、提出了单像素封闭边缘提取算法。该算法采用Canny算法获取的基础边缘与基础边缘端点灰度等高线相融合的方式产生具有单像素特征的连续封闭边缘,实验结果表明,该算法可有效地实现对灰度变化复杂图像进行封闭边缘提取,可应用于获取灰度类似的封闭区域、边缘测量等方面。3、研究了基于SVM固定形态目标识别技术。该技术通过对固定形态目标采用支持向量机(SVM)对其多项图像特征进行训练,获取差异性特征子集及其特征参数形成匹配模板,并将此模板用于识别算法中。该技术应用在集成电路引脚平整度自动测量系统中,使得该系统的测量准确率达到93%以上,错误接受率为0。4、研究了基于实时学习模式的非固定形态目标识别技术。该技术采用封闭边缘算法获取局部封闭区域并提取这些区域的特征,然后通过系统与用户的交互实现实时学习过程,根据用户的选择形成分类法则。该技术可用在金相组织含量分析系统中,实验表明,该系统定量分析误差仅为±1%,且适用范围更广。更多还原

【Abstract】 With the combination of optical microscope and computer technology, the development trend of optical microscope is automation and intelligentizing which greatly enhance its performance and effectiveness and has an important significance to optical microscope’s development and application.According to the future development tendency of optical microscope and the characteristics of microscope’s observation,the observed object can be divided into two categories: fixed shape object and non-fixed shape object. The key technologies of image recognition about two categories object are studied in this dissertation. The main contents of this dissertation includes: a multi-focus image fusion algorithm which can extend the high power microscope’s depth of field, an single pixel wide enclosing image edge extraction algorithm, an automatic identification and measurement algorithm for fixed shape microscopic image objective and a classification algorithm base on real-time learning pattern which can be applied to identify non-fixed shape microscopic image objective. The effectiveness and applicability of these algorithms was validated through the contrast and performance experiments.The major innovations of the dissertation are summarized as follows:1. A multi-focus image fusion algorithm based on noise adaptive filter and the nearest neighbor’s weight is developed to extend the high power microscope’s depth of field. This algorithm can effectively reduce the influence of the noise, and has very good performance in different magnifying power objective lens.2. An algorithm based on Canny and grayscale contour line to extract single pixel wide enclosing image edge is proposed. Acquiring the initial edge which have single edge effect by using Canny algorithm. The initial edge’s threshold was automatically calculated to reduce false edge and obtain the basic edge for next steps. According to the grayscale neighborhood of the basic edge’s end points, the gray value of grayscale contour line was calculated. On the fusion condition of basic edge and grayscale contour line, the closing edge can be create from the endpoints of basic edge. Experimental results indicate that this algorithm can generate enclosing edge efficaciously which can be used to obtain enclosing region and edge measurement.3. Technology of fixed shape object recognition base on SVM-based is studied and realized. According to the fixed shape of objects, using the SVM to training the variety of features to obtain key feature subset and feature parameters for automatic detection’s matching template. With the application of matching template to the flatness automatic detection system for Integrated Circuit pins, the accuracy is over 93% and false-acceptance error rate is 0.4. Technology of non-fixed shape object recognition base on real-time learning pattern is studied and realized. The analysis system extracted the features of region which surrounded by enclosing edge. Through the human-computer interaction, it learned the definition of specific enclosing region selected by the user to form classification rule about features and according to these rules to classify the other region. With the application of the metallographic structure contents qualitative analysis system, experimental results indicate that the measurement error of the polyphase structure’s contents about metallographic is less than 1% and suitable for different shape metallographic structure.更多还原