【作者】 邓泽峰；

【导师】 熊有伦；

【作者基本信息】 华中科技大学 ， 机械电子工程， 2007， 博士

【摘要】 图像复原是利用退化现象的某种先验知识(退化模型),按退化的逆过程重建图像的技术。噪声干扰和运动模糊是工程中常见的两种退化类型,本文研究了噪声滤波和运动模糊参数识别算法,以及这些算法在RFID倒装设备飞行视觉系统中的应用。椒盐噪声严重降低了图像质量,其滤波性能的好坏直接影响后续图像处理的结果。本文提出了一个基于数学形态学的高污染椒盐噪声滤波算法。首先采用开、闭运算检测图像中的噪声并分类;然后针对检测到的椒噪声和盐噪声,设计相应的形态学滤波器;对滤波图像中的黑斑和白斑则采用一个简单的智能斑点擦除算法;最后加权求和形成开闭序列(OCS)滤波算法。仿真结果表明,与其他算法相比,该滤波算法能够有效地检测和滤除图像噪声,并保留更多的图像细节。深入研究了图像匀速直线运动模糊的点扩展函数,提出其关键参数(模糊方向和模糊长度)识别的频域方法和微分方法。频域方法:通过分析模糊图像频谱图中的暗条纹特性,揭示了模糊方向与其垂直的关系,采用基于Radon变换的极大值(MRT)识别模糊方向;通过实验建立了模糊长度与暗条纹间距的反比例数学模型(IPM),并由运动方向上的Radon变换测定暗条纹间距。微分方法:采用双线性插值技术计算方向微分,模糊方向对应于图像最小方向微分(MDD);定义水平方向上微分图像的自相关,由微分自相关间距(PDA)计算模糊长度。研究了噪声对频域方法和微分方法的影响,并比较了两个方法的性能。实验表明,频域方法计算简单,结果准确,但有噪声污染时效果较差;微分方法不但计算准确,还具有一定的抗噪声能力。设计了面向RFID装备的飞行视觉系统,所提出的OCS滤波算法、模糊参数识别的频域方法和微分方法都在其中得到应用。依据识别参数利用维纳滤波复原图像,并进行了模板匹配实验。应用实例表明,本文的图像复原技术可有效改善图像质量,提高匹配的准确度和精度。更多还原

【Abstract】 Image restoration is an image reconstruction technology that image is rebuilt with some prior knowledge (degraded model) according to counter-process of degeneration. This dissertation involves the specific topics on the two models of degradation: noise jamming and motion blurring. The algorithms of noise removal and parameter identification for motion-blur are further studied and tested on the on-the-fly vision system of RFID flip chip bonder.Impulse noise can seriously deteriorate image quality and the performance of its filter affects directly the result of subsequent image proceeding. A novel filter based on mathematical morphology for high probability impulse noise removal is presented. Firstly, an impulse noise detector using mathematical residues is proposed to identify pixels which are contaminated by the salt or pepper noise. Then the image is restored using specialized open-close sequence algorithms that apply only to the noise pixels. Finally, black and white blocks which degrade the quality of the image will be recovered by a smart block erase method. Experimental results demonstrate that the proposed filter outperforms a number of existing algorithms and can remove most of the noises effectively while preserving image details very well.The point spread function for the image blur of uniform linear motion has been explored. Two methods based on frequency-domain and directional derivatives are proposed to identify two major blur parameters of the point spread function– blur direction and blur length. In the first method, the reason for the occurrence of black strips in the spectrum image is analyzed and the black strips that are perpendicular to the motion-blur direction are detected using radon transformation. The mathematical model of the relationship between blur length and the pitch of the dark lines is estimated based on the curve fitting algorithm. In the second method, the directional derivative is defined and the directional sub-pixel is calculated through bilinear interpolation. The angle corresponding to the minimum global directional derivatives is identified as blur direction. And the blur length is estimated by the minimal value of the derivative autocorrelation. Experimental results show that the frequency-domain method facilitates real-time calculation, while is not suitable for the noise image. The derivative method not only achieves accurate results, but also has strong noise immunity.A novel on-the-fly vision system applied to RFID flip-chip bonder has been presented. The feasibility and availability of the proposed algorithms above have been tested and verified on RFID devices. Based on the identified parameters, the Wiener filtering is carried out to recover the motion-blurred image. And the restored images are further analyzed by pattern matching. Examples of application show that the proposed algorithms can efficiently enhance image quality and also improve the accuracy and precision of pattern matching.更多还原