基于在线度量学习的大规模图像检索技术研究

With the rapid development of digital cameras and social media networks, the rapid growth of the size of the image data on the Internet, how to find the required image data from a large-scale image quickly and efficiently has become the urgent needs of the user. Thus, image retrieval technology is becoming an important technique for many real-world multimedia applications. However, large-scale image retrieval has the low learning efficiency for high-dimensional sparse feature and the low retrieval efficiency for cross-view feature. We study the large-scale image retrieval method by combing online learning, metric learning, and deep learning. The main contents of our research are summarized as follows. 1) We propose a online metric learning based visual codebook construction method to perform feature selection for high-dimensional visual feature in order to obtain the semantically related Bag-of-Words model; 2) We propose a sparse online learning of image similarity method to improve the learning efficiency and similarity calculation efficiency of large-scale high-dimensional visual feature; 3) We propose a deep learning based cross-view feature fusion method to represent different features in a shared space, then exploits online metric learning to measure the image similarity, which aims to improve the retrieval performance for cross-view features.

随着数码相机和网络社交媒体的飞速发展，互联网上图像数据规模急剧增长，如何快速有效地从大规模图像数据中找到所需要的图像已成为用户的迫切需求。因而，图像检索成为许多现实世界多媒体应用的重要技术。但目前大规模图像检索存在高维稀疏特征的学习效率低、跨视图特征的检索性能低等问题。本项目拟结合在线学习、度量学习、深度学习对大规模图像检索展开研究，主要内容包括：1）提出一种基于在线度量学习的视觉词典构建方法，对高维视觉特征数据进行特征选择，得到语义相关的词袋模型；2）提出一种稀疏在线图像相似性度量学习方法，提高大规模高维视觉特征的学习效率和相似度计算效率；3）提出一种基于深度学习的跨视图特征融合方法，将在共享空间中不同类型的特征映射为相应的特征向量，通过在线度量学习算法进行相似性度量，以提高跨视图特征之间的检索性能。

本项目针对大规模图像检索存在高维稀疏特征的学习效率低、跨视图特征的检索性能低等问题，结合在线学习、度量学习、深度学习对大规模图像检索展开研究，主要内容包括：1）提出一种基于在线度量学习的视觉词典构建方法，对高维视觉特征数据进行特征选择，得到语义相关的词袋模型；2）提出一种稀疏在线图像相似性度量学习方法，提高大规模高维视觉特征的学习效率和相似度计算效率；3）提出一种基于深度学习的跨视图特征融合方法，将在共享空间中不同类型的特征映射为相应的特征向量，通过在线度量学习算法进行相似性度量，以提高跨视图特征之间的检索性能。本项目已在IEEE TIP、TMM、ACM TOMM、ICIP等国际期刊和会议上发表论文10篇，培养博士生3名、硕士生6名。