【作者】 陈国波；

【导师】 熊四昌；

【作者基本信息】 浙江工业大学 ， 机械电子工程， 2007， 硕士

【摘要】 基于体视显微镜(Stereo light microscope，SLM)的显微立体视觉系统已用于微操作、微装配等领域，作用之一是通过视觉反馈实现二维或三维的高精度自动定位，引导机械手完成指定的操作；作用之二是挖掘立体视觉中包含的三维深度信息，用于三维信息的测量。本文研究了SLM显微立体视觉的成像机理、图像预处理、图像配准和利用聚焦评价函数进行深度恢复等主要内容。研究的结果可用于小尺度对象的三维信息测量和微操作、微装配中的三维定位。本文研究的内容是使用单目显微镜进行物体表面的三维重构。由于显微镜景深有限，无法通过一幅图像得到完整清晰的物体表面信息，所以通过物体在Z坐标上的移动，显微镜对物体表面进行层层扫描，得到显微序列图像，同时得到每幅图像初步的Z坐标值；然后提取出每幅图像中的清晰区域，融合成一幅图像，得到一幅全聚焦图像；另一方面，将序列图像中的每幅图像清晰部分进行聚焦测度计算，通过对聚焦测度计算结果进行高斯插值得到每一点的深度；最后使用深度信息对显微样本进行三维重构，得到显微样本的三维模型。本文采用了显著边界特征匹配法，对图像进行自动配准，经过分析和实验表明此配准方法能够达到高精度图像融合的要求；其次，分析了小波算法在图像分解与重构中的原理及其融合规则，并在此基础上采用了一种全新的基于能量规则的小波融合序列图像的方法，在实验中表明了小波能量融合的方法能相当好的抑制图像融合失真问题；然后，本文使用计算图像熵的方法对图像清晰度进行评定，效果良好，同时加以均值和标准差两种评价标准对本文所有的融合实验结果进行了综合的评价，避免单一参量的片面性；最后，采用高斯插值算法来估计精确聚焦位置，从而得到显微样本的物体表面深度信息，进行图像的三维重构。本文提出了根据提取聚焦部分图像的小波参数，分析聚焦部分和模糊区域在小波变换域内图像能量分布密度的差异，以其高频部分与低频部分的比值为特征参照值，取其最大值为新的小波系数的融合规则。此方法能很好的判断、剔除成像中散焦部分及解决一些融合中的失真问题，比常规的基于区域的以像素点取最大值的小波融合方法有明显的优势。根据单目显微三维重构的特性，采用了高斯插值算法来估计精确聚焦位置，从而得到显微样本的物体表面上每一点的深度信息，得到目标物体的三维模型。更多还原

【Abstract】 Micro stereovision based on stereo light microscope (SLM) is used insome Micro-domains such as micromanipulation, microassembly, etc.There are mainly two purposes of micro stereovision: the first is that it isused to auto position in micromanipulation and microassembly by thevision feedback and helps the mechanical hand to complete-the expectedoperation; the second is that it contains the 3D depth information which isused in the measurement of 3D attributes. In this paper, some importantissues correspondent with micro stereovision based on SLM are studied forsolving problems of micro positioning and micro measurement, such as themathematical description of the imaging process of SLM, the imagepreviously treatment, the image matching algorithm and a depth estimationalgorithm interpolates a small number of focus measure values to obtainaccurate depth estimates etc. Results derived from this paper can be used to3D measurements of objects with a small scale and 3D positioning inmicromanipulation, and microassembly.The main content of this paper is reconstructs the object surface to 3Dsurface use one-eye microscope. We can’t obtain the clear and allinformation from one image because of the limitation of the microscopicfocal length. We displace the object in the Z axial so we can obtain a sequence of object images, then we pick-up the focus area in every imageand fusion them in a image. The sum-modified-Laplacian (SML) operatoris developed to provide local measures of the quality of image focus. Adepth estimation algorithm interpolates a small number of focus measurevalues to obtain accurate depth estimates information to reconstruct theobject surface to 3D surface.Firstly, the paper adopts the edge characteristic image registration, andthe practical experiments prove this auto-matching method is effective.Secondly, the research analyses the principle of image decomposition andreconstruction based on wavelet-pyramid method. From the former way thepaper proposes a new wavelet transform algorithm based on the energyfusion rules. And the new way can remove the defocused areas showed bythe experimental result. And through the entropy and others evaluation, thecomprehensive evaluation result was achieved. Thirdly, in order to obtain3D structures of the object to reconstruction the object surface to 3D image,Gaussian interpolation algorithm has been selected to interpolate a smallnumber of focus measure values to obtain accurate depth estimates.Based on the different energy distributions of the low-frequencyband and the high-frequency band, which are achieved by discrete waveletstransform, an original method for wavelet coefficient combination isproposed in this paper. It takes the ratio of the two bands as a characteristicvalue and the maximum value is chosen as an integration rule for determining new wavelet coefficients. This method can remove thedefocused areas and reduce the distortion of the fused images. Finally, theadvantage of the proposed fusion approach is demonstrated clearly bypractical experiments, by comparison with conventional area based pixelselection method. Because of the characteristic of acquire images use onesensor, Gaussian interpolation algorithm has been selected to interpolate asmall number of focus measure values to obtain accurate depth estimates to3D reconstruction.更多还原