基于深度神经网络的哈希学习及其在图像检索中的应用研究

In the context of large scale data, hash learning based method is promising for efficient image retrieval. However, the learned compact binary codes are still encountered by the problems of low retrieval precision and hard to lean the hash functions under the binary constraint. This project aims at improving the retrieval performance of compact binary codes by focusing on the two key issues of accurately exploiting and preserving semantic similarity and optimizing hash functions. To fulfill this goal, the project studies the deep learning based hashing methods and the strategies of similarity preservation and its application in the multimedia retrieval. In particular, this project studies the following two aspects: deep learning based discrepancy minimization hashing, multiple output deep learning based hashing for multiple label image retrieval. This project will produce independent innovation achievements with international academic impact and lay the technical foundation for the fast retrieval of large scale image data.

基于哈希学习的图像检索在大规模数据环境下具有广阔的应用前景，但紧致的二值特征在用于检索时仍存在近邻检索精度偏低、难以精确学习等不足。本项目将从二值特征的对象语义相似性挖掘与保持、哈希映射的优化求解两个关键问题入手，主要研究基于深度神经网络的哈希学习方法和相似性保持策略及其在图像检索中的应用，着力提高二值特征的检索性能。项目具体研究内容将从基于深度神经网络的残差最小化哈希学习、面向多类标图像检索的多输出深度哈希学习两个方面开展。本项目将形成具有一定国际学术影响力的自主创新成果，为大规模图像数据下的快速检索提供技术支持。

在基金项目的资助下，项目组围绕深度神经网络模型用于大规模图像和视频检索的哈希学习开展研究。提出了两种用于将视觉数据编码为二值特征的哈希学习方法，能够在语义信息和近邻关系上进行有效刻画和编码，提高了哈希学习方法在大规模数据下的检索性能。基于研究成果，项目在顶级国际学术会议ICCV和重要国际期刊IEEE TMM上发表和接收学术论文2篇。