基于变化人脸的三维人脸表情识别.docVIP

 基于变化人脸的三维人脸表情识别# 李小利，阮秋琦** （北京交通大学信息科学研究所，北京 100044） 5 10 15 20 25 30 摘要：自动的人脸表情识别目前还是一个很具有挑战性的课题，特别是在对三维空间数据的 处理上。本文针对自动实现三维人脸表情识别提出了一种分析策略。该策略中最重要的一步 就是变化人脸的形成。为了获得变化人脸，首先网孔结构的人脸模型要转换为方格结构的模 型；然后，这些方格结构的模型被用于获取深度图并被标准化；最终，变化人脸就由标准化 的深度图得到。整个获取变化人脸的过程都是全自动实现的，而且后续的人脸表情识别的操 作也是完全不需要人工干预的。且为了验证整个策略的可用性，本文提取傅立叶谱作为分类 特征，并最终以高斯核函数的 SVM 获得最高的识别率为 85.33%，这一结果与先前的工作相 比具有很好的可比性。而且本文提出的方法最重要的一点是整个三维人脸表情识别的过程是 可以完全自动实现的，这对识别系统具有很重要的意义。 关键词：三维人脸表情识别；自动实现；深度图；变化人脸；傅立叶变化 中图分类号：TP391.4 3D facial expression recognition based on variation faces LI Xiaoli, RUAN Qiuqi (Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China) Abstract: Automatic facial expression recognition is recently a challenging problem especially in 3D space. This paper proposes a synthetic strategy for automatic 3D facial expression recognition. The most important part of this strategy is the generation of the variation faces. To obtain the variation faces, firstly, the 3D mesh facial models are converted into 3D grid models; secondly, 3D grid models are used to get the depth maps which will be normalized then; finally, the variation faces are obtained based on the normalized depth maps. The flow of variation faces’ generation is automatically implemented, and the following facial expression recognition is accomplished totally without any manual intervention, too. To validate this strategy, the Fourier spectrum feature is explored and its highest recognition rate is 85.33% using Gaussian kernelled SVM which represents to be comparative to most primary work. The most important thing of the proposed strategy is that it enables the whole recognition system to be implemented automatically which opens a promising direction for 3D facial expression recognition. Key words: 3D facial expression recognition; automatic implement; depth map; variation face; Fourier transform 35 0 引言 人脸表情识别的研究已历经了几十年，并在以二维静态图像以及二维图像序列[1，2]为 基础的研究上取得了很大的成果。目前，此课题又将研究方法定位为对三维人脸表情的识别， 以及怎样在保证不降低识别效果的前提下实现自