【作者】 孙永宣；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2013， 博士

【摘要】 随着互联网的普及和使用，特别是互联网上的图像和视频数据成指数增长的趋势，如何对这些数据进行处理分析和学习，从中得到有用的信息和知识，对完成基本的视觉任务，满足用户需求非常重要。集成学习作为近些年来一种有效的机器学习方法在各个领域都得到了广泛的应用，特别是在机器视觉领域成为了研究的热点。论文在有侧重地总结集成学习在计算机视觉任务中的图像表示、图像分割和图像分类三个重要方面的应用研究现状以及集成学习发展现状和理论的基础上，重点研究了集成学习在这三个视觉任务中的应用。论文通过分析三种视觉任务存在的问题，在视觉任务的图像特征表示、图像分割标记以及图像分类上，通过集成学习方式，有效的提高了视觉任务性能。论文取得的成果和主要创新点包括：（1）针对图像的特征表示问题，对无标记数据集原始数据分布，提出独立子空间中的特征增量学习方法，得到结构化的特征基元矩阵，形成有效的特征空间表达；同时提出基于AP聚类的样本距离度量方法，并定义样本奇异性，可有效检测原始数据空间分布下的奇异点，实现样本选择。该方法既可用于图像分类，也可用于图像检索。实验结果表明，所提方法可有效发现奇异样本，优于当前主流特征表示与学习方法，同时，也验证了奇异性图像可显著提升分类准确率的结论。此外，本文还对多核多特征集成进行了研究，实验结果表明了多核多特征的集成在场景分类上更有效。（2）研究了无监督集成聚类下的图像分割，针对同一种单聚类器对不同问题有不同的表现，不同的单聚类器对同一个问题的分析性能各异问题，采用集成聚类有效提升单聚类器精度和泛化能力。图像分割问题定义为像素级的标记问题，将集成聚类的算法延伸到图像分割处理领域，提出一种基于集成聚类机制的图像分割方法。通过集成多基聚类器的图像分割结果，对产生的多个差异性的基聚类器成员进行对齐，使用加权投票机制，设计共识函数，合并集成分割结果。在UC Berkeley图像库上实验结果表明，集成聚类比单聚类器分割结果从定性上更符合人类视觉感知，定量上也有较大的改进。（3）研究了迁移集成机制下的图像分类。采用迁移学习策略，在相关领域或任务间设计共享知识的学习方法，从一个或多个源域任务中抽取知识并将其应用于解决目标域任务，通过再利用现有历史训练数据，实现知识在不同领域、任务和分布间的传播，提升对小样本及训练、测试样本分布不同的学习问题求解的泛化性能。着重研究如何使用迁移学习理论解决小样本和训练、测试样本不同分布条件下的图像分类任务，通过考虑目标域、源域样本分布之间的差异性，提出了对原始多源集成迁移学习算法中的损失函数引入了协变量偏移修正；进一步，通过在每轮迭代中计算源域的可迁移度，进行源域选择并滤除不相关源域，一定程度上抑制了多源域条件下的“负迁移”现象，降低了计算开销。实验结果表明，新方法可有效实现小样本条件下的图像分类任务。更多还原

【Abstract】 With the development and popularization of Internet technology, the data of images and videoshave increased exponentially. The valid inference and learning algorithms for the computer visiontasks are very important, which would help people to obtain useful information and knowledgeconveniently. Recently, ensemble learning, an effective machine learning method, has been widelyused in various fields and attracted increasing attention especially in computer vision. The state ofthe art of ensemble learning are summarized in this paper. We mainly focus on the three applicationswhich include image representation, segmentation and classification. By analyzing the existingproblems involved in the above tasks, we explore ensemble learning method to improve theperformance of feature representation, segmentation label and object classification. The mainachievements and innovations can be concluded as follows:(1) We focus on the problem of image feature representation. A feature incremental learningmethod in independent subspace in the original unlabeled data space is proposed to get structuralfeature element matrix and form effective feature space representation. At the same time, a distancemeasuring method based on AP clustering is provided and the outlier of sample is defined which candetect outlier in the original data space to help the sample selection. This method can be used notonly in image classification, but also in image retrieval. The experiments show that our method canfind outliers effectively and achieve better performance than other popular feature representation,learning methods, and classification. Meanwhile, this paper also study multi-kernel and multi-featureensemble. The experiments show that it can get more effective results on scene classification.(2) Unsupervised clustering ensemble for image segmentation is proposed. As the same singlecluster will get different performance under different problems and the different clusters will getdifferent performance under the same problem, the clustering ensemble is used to improve theaccuracy and generalization of clusters. Image segmentation is defined as pixel level labelingproblem and clustering ensemble algorithm is applied to the task of image segmentation. Aclustering ensemble mechanism for image segmentation is provided. By integrating multi-clusterimage segmentation results, the differences of single cluster are aligned. Then, with weighted votingstrategy, the ensemble segmentation results are combined. The experiments on UC Berkeley imagedataset show that the segmentation results by ensemble clustering re consistent with humanperception and better than single cluster. The estimate indicators also have been improved greatly.(3) Classfier ensemble by transfer learning is studied to solve the problem of imageclassification. By design the transfer learning strategy, the knowledge is shared among related tasks.It can extract knowledge from one or multi-source domain to solve the problem in target domain.The knowledge in different domains, tasks and distributions can be broadcasted by reusing thetraining data. The generalization of computing can be improved when solving the problems uding few samples or the training and testing data set with different distributions. We focus on how to usetransfer learning method to solve the image classification under the above cases. Covariate shift isincorporated into the loss function to cope with the distribution differences between source domainand target domain. Additionally, the transferability of source domains is evaluated and eliminateirrelevant source domain gradually. Our method enhances the effectiveness in choosing availablesource domain, avoids negative transfer and promotes computational efficiency. Experiment resultsshow the proposed algorithm can achieve higher classification accuracy by using less training data.更多还原