面向图像分类的无监督迁移学习算法研究

Image classification is a fundmental vision task which recognizes the identification of the object and further provides the support for more complex vision applications. Thus, it is necessary to investigate image classification methods. However, traditional classification methods maybe fail because the coming or target-domian images are blessed with different data distributions from the old or heterogeneous source-domain images. If we want to learn classification model, amounts of labeled images are equipped for its stablity but this is so expensive in time and could not be allowed to manually labeled them in some situations. Thus, we utilize unsupervised transfer learning methods to make use of knowledge from source domain to address the above issue. Specifically, the proposal readily develops the new unsupervised transfer learning methods based on the distribution discrepancy metric, the geometrical structure within the dataset, and the utility of source labels, respectively, with hope of achieving the state-of-the-art classification performance and the wide applicabilty.

图像分类是计算机视觉应用中的一项基本任务，旨在识别图像中的物体，为其他复杂视觉应用提供支撑，因而研究图像分类方法非常有必要。但是，实际应用中传统分类方法可能会失效，这是因为现有数据或目标域数据分布不同于已有标注数据或源域数据分布。要想训练稳定的分类器就必须要标注大量的目标域数据标签，但其成本代价大，甚至有时无法标注样本。为此，本项目拟借助无监督迁移学习算法利用源域数据语义信息来克服上述问题。本项目具体从分布差异度量、捕获语义信息和合理使用源域标签这三个方面设计新的无监督迁移学习，旨在提升现有迁移方法的图像分类性能及其实用性。

图像分类是计算机视觉应用中的一项基本任务，旨在识别图像中的物体，为其他复杂视觉应用提供支撑，但是，实际应用中传统分类方法可能会失效，这是因为现有数据或目标域数据分布异于已有标注数据或源域数据分布。为提升模型的可迁移性，本项目从分布差异度量、全局语义关联、信息论的特征冗余、标签分布平衡性、通道混淆等方面设计了9个新型无监督迁移学习算法，分别从不同层面解决域间分布配准的问题，其中诸多算法在发表当年都能获得最佳性能，并且有些算法被应用到语义分割、目标跟踪、行人重识别等领域，极大地扩展了图像分类的应用范畴。