【作者】 孙建中；

【导师】 熊忠阳；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2013， 博士

【摘要】 在图像信息智能分析中，不确定性问题的研究成为一项重要而又具有挑战性的工作，概率图模型为解决这一问题提供了一种重要途径。本文对基于概率图模型的图像局部特征描述、非匀质图像分割和视觉跟踪的问题进行了深入研究。图像局部特征描述的核心问题是不变性(鲁棒性)和可区分性。然而特征描述子的可区分性的强弱和其不变性是矛盾的，也就是说，一个具有诸多不变性的特征描述子，它区分图像局部内容的能力就稍弱；而一个非常容易区分不同图像局部内容的特征描述子，它的鲁棒性往往比较低。因此，在基于局部特征的图像处理中，研究不仅具有较强不变性、还具有较好区分性的特征描述子有着重要意义。非匀质图像是包含有多个灰度级目标的图像。传统方法中的边界模型虽然能够分割非匀质图像，但对噪声、弱边缘或不连续边较敏感，而区域模型中分片匀质的统计假设，使其无法分割灰度非匀质图像。因此，研究基于概率图模型融合局部特征信息的区域模型，对非匀质图像进行快速分割的问题，有着较重要的学术价值和应用前景。视觉跟踪过程中，目标表观的动态变化容易造成目标漂移或丢失问题。传统的模型更新算法大多采用固定的先验模型克服该问题。一方面，普适的固定先验容易使被跟踪目标漂移到相似的对象；另一方面，受限的固定先验因无法处理突变的目标表观而导致跟踪失败。因此，在跟踪过程中，利用概率图模型解决不确定性信息的能力，构建在线自适应先验表观模型是具有良好应用前景的研究方向。论文取得的主要成果与创新工作概括如下：①在形状变化、被遮挡、噪声情况下，存在难以完整分割目标形状的问题，为此，提出一种采用隐含形状约束马尔可夫随机场模型的形状分割方法。引入目标的先验形状知识，用水平集符号距离函数隐含表示目标的先验形状模型；以先验形状模型作为约束构造出MRF能量函数；采用graph cut法求解能量函数极小值，利用形状对准和最大流法演化初始轮廓，快速准确地分割出目标的形状。实验表明新方法能有效快速分割带遮挡、噪声以及发生形状变化的目标，增强了形状分割的鲁棒性。②针对灰度非匀质图像分割困难及效率低下的问题，提出了一种基于马尔可夫随机场的局部区域活动轮廓模型快速分割方法。该方法结合核函数和割测度定义一个新的能量函数。一方面，在中心点被核函数掩模的局部区域内，用邻近点的加权均值拟合数据项，能有效处理图像的非匀质分布。另一方面，用割测度逼近的曲线长度作为全局正则性，利于轮廓快速定位于物体边界。在轮廓演化过程中，使用基于栅格图的最大流算法，避免了传统模型计算代价高昂的水平集函数。合成图像和真实图像的实验结果表明，提出的方法能有效快速地分割灰度非匀质图像中的弱边缘物体及复杂的多灰阶结构物体；对初始轮廓线位置和噪声具有较好的鲁棒性。③针对目标跟踪过程中，可变目标表观的特征数据会发生“分布漂移”的问题，基于时间Dirichlet过程混合模型，提出了一种学习多模表观模型的方法并用于目标跟踪。以时间Dirichlet过程为先验分布，把先前估计的目标样本划分为不同的聚集，使得每个聚集表示一类表观，每个表观类被建模为判别式分类器；基于贝叶斯后验推断，权衡先前表观模型的分类误差和拆分聚集的代价，从数据中自主学习表观模型；基于Noisy-OR模型，以贪心（Greedy）策略协同各表观分类器判别出目标。仿真结果表明该方法能较好的跟踪可变目标表观，改善了目标跟踪性能。④为有效解决视觉跟踪过程中的目标“漂移”问题，基于分层Dirichlet过程演化聚类模型，提出一种采用自适应先验表观模型的目标跟踪方法。该方法在一致的架构内融合HDP-EVO演化聚类模型和在线Boosting学习。以分层Dirichlet过程为先验分布，对总体表观示例进行聚类分析，获得随时间自适应演化的表观类先验知识；另一方面，利用共享的表观类混合比例的权重平滑约束各时刻的表观模型；改进Gibbs抽样过程，使之能融入目标示例的分类误差，并交替迭代地从数据中自主学习聚类和表观分类器；根据表观模型中各表观类的权重系数组合它们的分类评分去定位目标位置。仿真实验表明新方法学习的表观模型能较鲁棒的自适应于目标的表观变化，提高了目标跟踪精度。更多还原

【Abstract】 Information uncertainty in image processing has become a new hot-spot as achallenging research. Probabilistic graphical model provides an important means forresolving the uncertainty of intelligent information field. This paper made an intensivestudy on local features of an image, segmentation for the in-homogenous images andvisual tracking based upon probabilistic graphical model.Robustness and discrimination are the core problems of different kinds of localfeature descriptors, but robustness and discrimination are contradictory. Namely, for afeature descriptor with robustness, its power of discriminating the content of an image isweak. Correspondingly, the one with discrimination has worse robustness. Hence, tobalance the trade-off between robustness and discrimination, the research based on localfeature descriptors has very important value in mage processing.Image with intensity in-homogeneity consists of multiple gray levels. Althoughclassical edge-based models can segment the image with intensity in-homogeneity, theyare sensitive for segmenting the objects with weak or discontinuous boundary and noise;region-based models have been successfully used in binary phase segmentation with theassumption that each image region is statistically homogeneous. However, region-basedmodels do not work well for the image with intensity in-homogeneity. Hence, theresearch of region-based models about integrating local feature information based onprobabilistic graphical model has very important academic and application role for thein-homogeneous image.Appearance variations of the target object will cause the drifting or missingproblem of the target object during tracking. To cope with the problem, most existingonline methods utilize fixed prior model to update appearance model. These methodsare either too generic to drift to the similar objects or too restrictive to fail in dramaticchanges. Therefore, a promising research direction is to construct an adaptive priormodel which can adapt to changes incrementally during tracking.The main contributions of this paper are summarized as follows:First of all, to deal with shape segmentation of the object in complicated scenes, anovel method based on MRF model with implicit shape prior was proposed. First, theobject was segmented using the initial contour. Then, an implicit shape model with levelset signed distance function was built to improve the accuracy of object segmentation, and a MRF energy function with the constraint of the existing shape prior wasconstructed. The optimal value of energy function could be found by graph cut method.The precise segmentation of object was obtained by applying the shape alignment andmax-flow method to evolve the initial contour. Experimental results show that theproposed method can better cope with the clutter and noise, as well as partial occlusionsand affine transformation of the shape. Thus the robust stability of segmentation resultis improved.Secondly, to solve effectively the difficult and ineffective segmentation for thein-homogenous images, a novel fast method based on the local region active contourmodel was proposed. A new energy function was defined by combining kernel functionand cut metric. Utilization of the kernel function was favor of computing thein-homogenous distribution of local regions effectively. On the other hand, betterapproximation of the curve length by cut metric could help the contours to evolve intothe object boundary quickly. In addition, in the evolving process of contours, amax-flow method was adopted, which avoided an expensive computational level setmethod. Experimental results using synthetic and real images show that the proposedmethod can effectively segment objects with the weak boundary in in-homogenousimages, as well as the complex structure objects with multi-gray levels. At the sametime, the method is robust to noise and the initial contours.Thirdly, to cope with appearance variations of the target object during visualtracking, a nonparametric Bayesian multi-modal appearance model for learning overtime was proposed. First, by taking the temporal Dirichlet process as prior distribution,the proposed model separated target samples previously estimated into several clusters.Each cluster represented a certain type of the target appearance, which was modeled asdiscriminative classifiers. Then, to balance the trade-off between the classification errorof appearance model and the cost for splitting the clusters, the multi-modal appearancemodel was automatically learned by the use of Bayesian posterior inference. Finally,based on the Noisy-OR model, a greedy algorithm was used to discriminate the targetobject by combining the outputs of appearance classifiers. The simulation results showthat the proposed method can robustly track an object under rapid appearance changesand improve the tracking performance.Fourthly, to effectively solve the drifting problem of a varying target object duringtarget tracking, an adaptive prior appearance model was presented. First, the methodcombined hierarchical dirichlet process evolutionary clustering model and online boosting learning into a coherent framework. By taking the hierarchical dirichletprocess as prior distribution, the prior appearance knowledge could adapt to changeover time. On the other hand, appearance model of each moment was smoothlyconstrained by the mixture proportion of a type of appearance cluster. Then, to balancethe classification error of appearance model and the cost for splitting the clusters, themulti-modal appearance model was automatically learned by the use of Bayesianposterior inference. Finally, based on the weighting factor of appearance clusters, thetarget object was discriminated by combining the outputs of appearance classifiers. Thesimulation results show that the learned appearance model can adapt to the appearancevariations of the target object and achieve better tracking results with high accuracy.更多还原