高性能低比特视觉搜索及芯片结构研究

项目名称： 高性能低比特视觉搜索及芯片结构研究

项目编号： No.61602011

项目类型： 青年科学基金项目

立项/批准年度： 2017

项目学科： 计算机科学学科

项目作者： 祝闯

作者单位： 北京大学

项目金额： 14万元

中文摘要： 直接利用数字图像这种更加直观的信息载体，通过智能终端进行视觉搜索将成为互联网信息检索的重要手段。然而实际应用中往往带宽和存储资源受限，所以研究高性能的低比特视觉搜索具有很重要的意义。围绕低比特视觉搜索这个问题，本课题从以下四个方面展开深入研究：（1）基于图像预处理的高鲁棒性特征提取（2）自适应局部特征选择（3）高性能视觉搜索算法（4）紧凑视觉特征提取硬件加速。通过自适应图像预处理算法，去除不同光照的差异，进而提取出对光线更加鲁棒的局部特征；通过融合局部特征自身的显著度以及先验概率信息，自适应选择更加具有辨别力的局部特征；通过多种查询扩展策略，提升视觉搜索的准确率和召回率；通过合理的软硬件分区和多层次混合流水线设计，对紧凑视觉特征进行高能效加速。本课题预期通过上述四方面的研究，从视觉搜索准确率和实时响应两方面改善低比特视觉搜索体验。

中文关键词： 视觉搜索；预处理；特征提取；特征选择；硬件加速

英文摘要： Conducting visual search through the smart terminal by using the more intuitive information carrier, such as images, will become an important manner for internet information retrieval. However, in practical applications bandwidth and storage resources are often limited, thus doing research on high performance low bitrate visual search is very important. In our project, we will conduct in-depth study for this topic from four aspects: (1) high robustness feature extraction based on image pre-processing (2) adaptive local feature selection (3) high performance visual search algorithm (4) hardware acceleration for compact visual feature extraction. We will remove the illumination differences for images through adaptive image pre-processing algorithm, and then extract more robust local feature. By combining the local feature saliency degree and the prior probability information, we will select more discriminative local features adaptively. Through a variety of query expansion strategies, we hope improve visual search precision rate and recall rate at the same time. Besides, we wish to achieve highly efficient hardware acceleration for compact visual feature extraction through reasonable hardware/software partitioning and multi-level hybrid pipelining design. In this project, we expect improving visual search accuracy and real-time response performance for low bitrate visual search experience through the study of the above four aspects.

英文关键词： Visual Search;Pre-Processing;Feature Extraction;Feature Selection;Hardware Acceleration