基于多边亲和度函数和序列图像的人脸超分辨率方法研究

项目名称： 基于多边亲和度函数和序列图像的人脸超分辨率方法研究

项目编号： No.61301183

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 周飞

作者单位： 清华大学

项目金额： 24万元

中文摘要： 视频监控在维护社会公共安全领域起着重要的作用。在监控视频中，人脸由于起到身份辨识的重要作用而备受关注。然而，在实际的监控视频当中，人脸所占区域分辨率很低，据此进行准确的身份辨识非常地困难。并且，目前的超分辨率技术仍然不足以解决该问题。为此，本项目提出基于多边亲合度函数和图像序列的人脸图像超分率方法，拟提高人脸图像或者视频中人脸区域的分辨率。当然，我们的目标不仅仅在于使得处理之后的人脸图像在视觉上更加清晰，更在于能够提高其辨别力。围绕着上述问题和目标，本项目主要研究的内容包括：基于多边亲和度函数的单帧人脸超分辨率方法、基于人脸约束多曲面拟合的序列图像超分辨率重建，以及结合利用训练数据和图像序列的方案等三部分内容。预期研究成果将提高人脸超分辨率乃至低分辨率人脸识别的理论研究水平及其在视频监控中的应用能力。

中文关键词： 超分辨率；多变亲和度函数；多曲面拟合；人脸猜想；

英文摘要： Video surveillance is of great importance in the field of the public security. In the surveillance videos, faces have attracted much attention due to their significant impacts on personal identification. However,in practice,facial regions are always of low-resolution. Based on such facial images, it is very difficult to make the accurate identifications. Moreover，existing methods of super-resolution are still inadequate to solve such problems. Therefore, we propose a face super-resolution method using multilateral affinity function and sequential images, with the aim of enhance the resolution of facial images or facial regions in videos. It is worthwhile to notice that our purposes are to improve the visual quality of faces as well as the discriminability of them. To achieve above purposes,the studies in this project include the following three contents: single-frame face super-resolution using multilateral affinity function, multi-frame super-resolution based on facial constrained multi-surface fitting, and the schemes to jointly exploit the training data and image sequence.The results of this project are expected to improve the theorical level as well as the application in video surveillance of face super-resolution and even low-resolution face recognition.

英文关键词： super-resolution；multilateral affinity function；multisurface fitting；face hallucination；