移动视觉搜索中基于特征保持的图像智能缩放技术研究

Recently, with the rapid development of communication technology and the popularity of various mobile terminals, mobile visual search (MVS) has become an urgent demand in various industries. At the same time, the image as the input of the interactive way, becomes a new way of interaction after the text and the voice, and is more and more favored by the majority of users. Nevertheless, due to wireless network with limited bandwidth, bandwidth fluctuation, transmission delay, as well as interoperability challenges brought by a variety of image retrieval systems, low bit-rate mobile visual search has attracted increasing attention of researchers, and has become a research hotspot in the field of multimedia information processing due to its wide range of applications in various industries. In order to satisfy the demands of low bit-rate and high interactivity, this project proposes a set of feature-preserving image resizing approaches for low bit-rate mobile visual search. As a result, the bottleneck appeared in the mobile terminal with large amount of data transmission, non-flexible feature extraction and interoperability difficulties in the variety of image retrieval systems will be resolved, which greatly supports the industrialization development of related technologies. In conclusion, this project has strong theoretical significance and practical application value.

近年来，随着通信技术的飞速发展和各种移动终端的普及，移动视觉搜索越来越成为各个行业领域的迫切需求，同时以图像作为输入的交互方式，成为文本、语音交互之后新型的交互方式，越来越被广大用户所青睐。由于无线网络环境的带宽有限、带宽波动、传输时延、以及多样化图像搜索系统互操作等方面带来的挑战，为提高用户体验，低比特率移动视觉搜索得到了越来越多研究者的关注，已成为当今多媒体信息处理领域的一项研究热点，并在各行各业中具有广泛的应用前景。为满足各项相关应用所需的低比特率和高交互性，本项目旨在提出一套基于特征保持的图像缩放方法以应用于低比特率移动视觉搜索，从而突破移动端所面临的数据传输量大、特征提取不灵活、多样化视觉搜索系统难以互操作的瓶颈，极大地支撑相关技术应用的产业化发展，具有很强的理论研究意义和现实应用价值。

本课题组自项目承担以来，经过四年努力，课题工作进展顺利，各项工作已完成，取得了较为丰硕的科研成果。具体地，针对数字图像中特征的过滤和选取与数字图像中特征显著性的计算，提出了一种基于语义分割和像素融合的图像智能适配显示方法、一种语义分割驱动的图像重构方法以及一种基于监控视频的运动管道提取方法。针对特征保持的图像自适应缩放与哈希编码保持的图像缩放，提出了一种深度自监督循环图像智能重适配方法和一种自适应特征保持的图像重适配方法。针对特征超分问题，提出了一种面向感知质量的立体图像超分辨率重构方法、一种视差感知的立体图像增强方法以及一种面向特征超分的生成对抗网络。相关科研成果陆续得到了国际著名期刊和学术会议的认可。共有30篇相关学术论文发表在国际著名的IEEE Transactions等国际期刊和学术会议上（全部标注本基金资助）。