大数据环境下基于社交网络的图像搜索技术研究

Due to the large scale of the image data, the research about perceptual hashing has its specificity and complexity, and high requirements on performance in addition. Therefore, existing image perceptual hashing algorithms could not cope with the requirements of large scale image retrieve in terms of robustness, discrimination, compactness, et al. . In conclusion, it is a challenging and difficult problem to research on perceptual hashing algorithms for large scale image retrieval. Under the big data environment, data has the properties of large scale, low value of density, and high requirement on efficiency. Due to the properities, existing image retrieval technologies are not suitable in big data environment. Our project focuses on image retrieval under big data environment, and intend to conduct theory and key technology research on the following four aspects: visual words hashing,social network orentied user encoding, efficient auto-encoding, and fast solution and optimization. Combining visual words and machine learning, we first research on feature hashing based on visual words. Then, combining high dimensional data-aware hashing with community detection, we propose the community-aware fingerprint coding method. By utilizing theory about multi-layer auto-encoder and transfer learning, we research on efficient data hashing algorithm. We also conduct solution and optimization algrithm, aiming for fast learning on large scale trainging data. Researches in our project not only enrich the theory and technology about image retrieval, but also promote subject development in relevant field. Besides, the research benefits data analysis and storage under big data environment, and since promote development on related industrialization.

由于大数据环境下的数据具有数据规模大、数据价值密度低、处理的时效性要高等特点，现有的图像检索技术在大数据环境中已不再适用。本课题将针对面向大数据环境的图像检索，从视觉词汇哈希、面向社交网络的用户编码、高效自动编码、快速求解优化等方面开展相关理论及关键技术研究。拟结合视觉词汇和机器学习的相关理论，研究基于视觉词汇的特征哈希；拟结合高维数据感知哈希（Hash）的思想和社区检测相关理论，研究具有社区感知的用户编码；利用多层自动编码和迁移学习，研究高效的数据哈希算法；研究求解优化算法，实现对较大规模训练数据进行快速学习的目的。本课题的研究不仅丰富了图像检索相关的理论和技术，而且还对相关领域的学科发展也起促进作用。更为重要的是，该课题的研究，有助于解决大数据环境下数据分析和存储的难题，将推动产业化的发展。

本课题针对面向大数据环境的图像检索，从视觉词汇哈希、面向社交网络的用户编码、高效自动编码、快速求解优化等方面开展相关理论及关键技术研究。首先结合视觉词汇和机器学习的相关理论，研究基于视觉词汇的特征哈希，通过局部保持和辨识力增强函数的优化，达到减少哈希过程中语义损失的目的；结合高维数据感知哈希（Hash）的思想和社区检测相关理论，研究了具有社区感知的用户编码；利用深度神经网络和哈希理论，研究了高效的数据编码方法，将学习好的模型参数迁移至哈希编码过程中然后构建浅层网络，便于数据的高效处理；研究了求解优化算法，利用MapReduce分布式并行计算框架实现对较大规模训练数据进行快速学习的目的。本课题的研究丰富了图像检索相关的理论和技术。项目资助发表SCI论文2篇，EI会议论文5篇， 其他期刊论文2篇。培养硕士生4名。项目投入经费20万元，支出16.0871万元，各项支出按预算安排执行。剩余经费3.9129万元，剩余经费计划用于本项目研究后续支出。