【作者】 艾春璐；

【导师】 陈莹；

【作者基本信息】 江南大学 ， 信号与信息处理， 2012， 硕士

【摘要】 图像匹配作为计算机视觉中的一个基本问题,在目标检测与识别、三维重建、模式识别、机器人视觉导航以及医学图像分析等诸多领域都有着广泛的应用。但在实际应用中,由于成像条件的变化,将造成图像的灰度失真和几何变形,这给图像的匹配问题带来了一定的困难。为了使图像的特征匹配技术达到更加完美的应用效果,国内外众多学者投身于图像匹配算法的研究之中,试图找到一种匹配精度高、鲁棒性强及实时性好的优秀算法。本文对图像的特征匹配问题进行了研究与探讨,并着重针对特征点的匹配问题展开了详细的研究,主要成果包括以下几个方面:1)针对复杂环境下特征点匹配精度、鲁棒性等下降的问题,在概率模型的框架下,提出了一种基于均衡化概率模型的特征匹配算法。首先利用重启动的随机游走(Random Walks with Restart, RWR)模型对每组候选匹配点的似然值进行估计,将其作为RWR模型中一组候选点与另外所有候选点之间的匹配概率值,其中,对RWR方法中的邻接矩阵进行均衡化分析和处理,得到一个平衡的邻接矩阵,这样做大大提高了匹配的准确性,最后再利用具有约束限制的时序方法从估计出的似然值中得到最优匹配集。2)在基于均衡化概率模型的特征匹配算法的基础上,实现了非刚性细胞形态的跟踪。首先利用基于均衡化概率模型的特征匹配算法实现目标整体定位,然后利用得到的特征匹配点以及一些简单的灰度、梯度信息进行目标的自动标定,并结合Growcut图像分割方法,实现噪声环境下非刚体目标边缘轮廓的精确跟踪。3)将基于均衡化概率模型的特征匹配算法应用于人脸跟踪中,实现了已知正面人脸标定点的情况下侧面人脸的精确跟踪,在提高跟踪结果准确率的同时也增强了算法的实时性。首先将提出的图像特征匹配算法应用到核岭回归(Kernel Ridge Regression, KRR)方法中,得到侧面人脸的标定点作为初始信息,同时提出了一种新的人脸跟踪算法,CAAM(Conditional Active Appearance Model)反向合成匹配算法,在假设已知正面人脸标定点的情况下,将原始的AAM(Active Appearance Model)反向合成匹配算法中形状模型与基本形状的对应关系,演变为侧面人脸特征点与正面人脸特征点之间的对应关系,通过建立形状模型,并根据反向合成匹配算法,对模型参数不断地迭代优化,最后得到精确的侧面人脸标定点。更多还原

【Abstract】 One basic problem in computer vision is image feature matching, which can be used widely in object recognition and analysis, 3D reconstruction, pattern recognition, robot vision navigation, medical image analysis and many other fields. But in the practical application, as the image conditions vary, which all can cause grey scale distortion and geometric deformation, these factors bring certain difficulties to image matching. In order to achieve perfect effect, scholars both at home and abroad devote themselves to researching the image matching algorithm, tring to find one algorithm which has high matching precision, strong robustness and fine real-time.In this article, the image feature matching problem is discussed and regarding the feature points matching problems the detail research has been done, the main achievements are as follows:1)A new algorithm of feature matching based on balanced probabilistic model is proposed in a probabilistic framework to solve the problem that feature point matching precision drops under complicated conditions. Firstly using Random Walks with Restart(RWR) the likelihood values of each candidate matching point sets are estimated, which are used as matching probability between a set of candidate points and other ones in RWR model. And a balancing analysis and treatment to the adjacency matrix of RWR is taken, then a balancing matrix will be obtained, which can improve the matching precision considerablly. Last an optimal matching set can be got by imposing a sequential method with mapping constraints in a simple way.2)Non-rigid cell contour tracking has been realized on the basis of feature matching algorithm which is based on balancing probabilistic model. Firstly the position tracking is accomplished using the feature matching algorithm which is based on balancing probabilistic model, then an automatic calibration to object is presented with the results of feature matching and some simple information about gray and grads. Last a precise object contour tracking under noisy conditions is presented accuately combined with Growcut, a kind of mage segmentation method.3)Profile face tracking accurately in the condition of knowing the frontal face can be achieved by applying the feature matching algorithm which is based on balancing probabilistic model, and it improves the accuracy and real-timing. This article applies the image feature matching algorithm to KRR(Kernel Ridge Regression) then calibration points are obtained as initial information. At the same time, a new algorithm is proposed --CAAM(Conditional Active Appearance Model) inverse compositional algorithm. Assuming the calibration points of frontal face is known, transforming the corresponding relationship between the shape model and basic shape in the original AAM into the one between profile face feature points and frontal face feature points. The model parameters can be optimized iteratively according to establishing shape model and inverse compositional algorithm, at last the precise profile face calibration points can be obtained.更多还原