【作者】 郭庆昌；

【导师】 王科俊；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 均值移动算法是一种迭代算法,在图像处理和计算机视觉中得到广泛的应用,但其理论分析仍然存在一定的不足。本课题完善了均值移动算法的理论证明,研究其在图像平滑、分割和目标跟踪中的应用。本论文完成的主要工作如下:针对以往文献对均值移动算法密度函数分析的不足,讨论了轮廓函数k(x)与以x为自变量的核函数k(?)之间的凹凸性关系,以及核函数k(?)与密度函数f(x)之间的凹凸性关系。比较分析了均值移动算法两种运动方式的迭代时间长短和步数多少。根据柯西收敛定理更好的证明了均值移动算法的收敛性。提出并证明了同窗内均值移动矢量间的夹角不大于90度等。总结了均值移动步长为牛顿步长、高斯-牛顿步长或其它非线性步长的条件;讨论了均值移动点在运动方向上取密度极大值的位置;分析了均值移动函数在其收敛点处密度一阶导数为零的充要条件等;研究了均值移动算法在图像平滑、分割和运动目标跟踪中的应用,针对图像平滑,提出了基于柱状HSV空间的均值移动图像平滑算法;分析了基于柱状HSV空间的多幅同场景图像的最优组合方法;通过实验比较分析了在不同颜色步长和空间步长情况下的图像平滑效果。针对图像分割,提出了根据当前帧图像边缘(由均值移动算法求出)和差分优化模板相结合的当前帧运动目标分割算法。针对目标跟踪,提出了双窗特征提取算法和基于像素分量的目标特征更新算法,以及跟踪窗自适应旋转和缩放算法。更多还原

【Abstract】 The mean shift algorithm is an iteration algorithm, which has been widely used in the image processing and the computer vision, still has some theoretical proficiency. Therefore, the theoretical proven of the mean shift algorithm is perfected as well as its application in the image smoothing, image segmentation and object tracking is studied in the paper. The main finished tasks in this paper are as fellows:In view of the deficiency in the analysis of the density function of the mean shift algorithm, the relation of the concavity and convexity between the profile function k(x) and the kernel function k(||x-c||~2), the kernel k(||x - c||~2) and the density function f(x) are discussed. Besides, the iterative step and time of two motion rules of the mean shift algorithm are compared. Some theorems are analyzed and proposed, such as the convergence of the mean shift algorithm is proved using the Cauchy theorem; the one that the angle of the mean shift vectors in the same window is less than 90 degree,etc. some conditions are summarized while the condition of the mean-shift step is Newton step, quasi-newton step or other step; the local maximum value of density in the moving direction is discussed; and the necessary and sufficient condition is finally analyzed that the first derivative of the density function is equal to zero on the convergence point, etc.At last, the application of the mean shift algorithm in image smoothing, image segmentation and object tracking is tested; the image smoothing algorithm based on the cylindrical HSV space is studied in virtue of image smoothing problems; the optimal combination methods based on the multiple same scene images in the cylindrical HSV space is proved; the effect of the image smoothing is discussed with different color and space steps. Considering the image segmentation problems, the object segmentation algorithm based on the image edges of the current frame (the edge is gotten based on the mean shift algorithm) and the optimal difference template is proposed. With a view to object tracking problems, the two-window feature extracting algorithm, the object feature updated algorithm based on the pixel, and the linear combination method of the tracking windows are proposed separately.更多还原