面向图像检索的无监督深度特征学习方法研究

随着多媒体数据的爆炸式增长，高效的分析、理解和检索这些数据成为一个紧迫而重要的研究课题。针对其中的检索问题，本项目拟以大规模图像数据为研究对象，重点研究符合人类视觉感知特点的语义特征提取与表示这一多媒体数据检索的基础问题。具体开展以下几个方面的研究：（1）基于生成对抗的端对端无监督检索框架设计；（2）基于无监督训练的融合语义信息和对几何变换鲁棒的特征学习方法；（3）面向深度特征的紧凑高效特征表示方法。最后，拟构建大规模语义图像检索示范验证系统，对项目的各项研究内容进行测试与评价。项目申请人及其团队在图像检索及深度学习领域工作基础扎实，经验丰富。本项目的研究可对本领域内非常重要的语义鸿沟问题的解决起到推动作用，为智慧医疗、平安城市、智能交通和网络搜索等实际应用提供理论支撑和技术支持。

With the explosive growth of multimedia data, analyzing, understanding and retrieval such data have become an urgent and an important task. For image retrieval, this proposal will take the large scale image data as the object to study, and mainly focus on the semantic feature extraction and representation in accord with human vision perception that are the fundamental problem of multimedia semantic understanding. We will carry out the project in the following aspects:　(1) Generative adversarial based end-to-end unsupervised learning framework for image retrieval; (2) Geometric transformation invariant feature learning integrated with semantic information based on unsupervised training; (3) Compact and efficient feature learning based on deep features. Finally, we will build a large scale semantic image retrieval system to evaluate the contents of the project. The principle investigator and his group members have finished lots of researches in image retrieval domain and deep learning domain. The execution of this proposal could have the positive effects on the bridging the semantic gap that is a most challenging problem in the this research field, and could provide some theoretical and technical support for intelligent medical, safety city, intelligent transport and web search.