目标检测在安全监控、身份认证等领域有广泛的应用。在训练样本不足的情况下，标准的检测器学习方法效果都会大幅下降。迁移学习方法可以通过从其他相关数据域迁移知识，来帮助提高检测器学习的效果。本文在可变形部件模型的基础上，研究应用于目标检测的迁移学习方法。可变形部件模型分别用根模板和部件模板描述目标的全局和局部表观特征。这些模板能作为可迁移知识，同时其可变形的结构特性也可以帮助提高迁移学习的效果。本文的主要贡献包括：1. 提出了基于可变形部件模型的模板自适应迁移学习算法。现有的用于目标检测的迁移学习方法在利用参考模板时未根据对象目标对参考模板进行参数调整，或只对全局的模板进行参数调整，限制了迁移效果。为了充分利用可变形部件模型的弹性结构帮助提升迁移效果，本文提出基于可变形部件模型的模板自适应算法。算法通过利用可变形部件模型中部件模板的隐变量获取更加准确的局部特征从而更好地调整模板参数。实验证明提出的模板自适应迁移学习算法优于现有的其他迁移学习算法。2. 为了迁移学习能获得合适的参考模型帮助提升学习效果,提出了基于特征对齐和滤波的隐变量聚类算法。当辅助迁移学习的数据域中因表观多样性存在多子类时，需要先对数据域进行聚类分析和知识选择才能使迁移学习获得对学习有帮助的知识，从而保证迁移学习的效果。通过基于隐变量的特征对齐和滤波，样本间的相似性可以得到更准确的度量。无需进一步的人工标定，算法也可以获得合适的聚类结果，并学习多组分可变形模型作为参考模型。实验证明了隐变量聚类算法的有效性和该算法对迁移学习带来的帮助。3. 提出了登录式目标检测框架。在很多实际应用中，用户只希望检测自己感兴趣的特定目标类，但在用户登录相关信息（如少量该类样本）前，检测系统无法得知目标种类。我们称该情况为登录式目标检测。本文提出的登录式检测框架，可以利用少量的登录样本和已有的参考模型，提供适用于登录目标种类的检测器。因为登录目标种类和已有的参考模型代表的种类存在不同的匹配可能情况，本文提出了决策选择算法和模型选择算法，帮助选择正确的模型生成算法，合理地利用知识使得能够学习得到适用于检测对象目标的模型。实验结果表明提出的登录式检测框架的检测性能优于已有的学习和检测框架。
Object detection is a hot topic in computer version which has been widely applied in many fields such as auto-focus and security monitoring. When lacking of sufficient training samples, the performance of standard detector learning approaches will drop dramatically. Such a demand of substantial training samples can be lessened by transfer relevant knowledge from other sources into the training procedure of the target object detector. This thesis aims to develop an efficient transfer learning technique for object detection based on the deformable part-based model (DPM). The DPM describes the global and local appearance features of the object by its root and part filters. These filters can be served as good prior knowledge and the deformable structure is fully exploited to improve the performance of transfer learning. The main contributions are:1. A filter adaption algorithm based on the deformable part-based model is proposed for transfer learning. Existing transfer learning approaches employ the auxiliary filters without adaption or only apply the adaption on rigid templates. To fully exploit the deformable configuration of the deformable part-based model and achieve higher performance gain,a DPM-based filter adaption algorithm for transfer learning is proposed. With the latent values of the flexible filters in the DPM, the local features can be captured more precisely so that the filters can be better adapted to the target object. Experimental results demonstrate the superiority of the proposed filter adaption method over the existing filter adaption approaches.2. A latent clustering method is proposed by feature alignment and filtering so that suitable auxiliary models can be generated for transfer learning. When there are many subcategories that have great appearance diversity in the offered source domain, the performance gain of transfer learning will be limited without knowledge categorization and selection. With the proposed latent clustering method, the similarity between the positive instances can be better measured and appropriate object subcategories can be obtained for training multiple component DPMs as auxiliary models, requiring no fine subcategory annotations. Experiments demonstrate the effectiveness of the proposed clustering method and its aid for transfer learning.3. A novel user-registered object detection framework is proposed. In many applications, users only want to detect specific target species that are undetermined to the detection system until the users register some relevant information like a few target samples. We call this scenario the detection of user-registered object. The proposed framework can generate an adaptive detector, form a limited number of user-registered target samples and several off-line trained auxiliary models. Experimental results demonstrate that the proposed framework can achieve superior detection performance to the existing detector learning and detection frameworks.
