基于LBP和LPQ的人脸表情识别

发布时间：2018-02-15 13:25

本文关键词： 人眼检测局部二值模式局部相位量化二维Gabor小波深度信念支持向量机　出处：《杭州电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：人脸表情是人类情感的载体,作为人类日常交往中一种重要的信息,反映了人类的心理状态。人们通过微小的表情变化,传递给他人不同的情感信息。随着科技的不断发展,人们对于人机交互的方式的要求更为多元化,如何让计算机读懂人脸表情信息成为当前一个热门的研究课题。如今表情识别技术受到了越来越多的重视,在诸多领域得到广泛地应用。本文对表情图像识别流程的三个步骤中用到的相关算法进行深入研究。研究内容以及创新性工作包括以下几点:(1)针对预处理环节切割纯表情图像问题,运用积分投影方向,定位人眼以及眼球精确区域。根据传统的“三庭五眼”面部模型切割出纯表情图像,用双线性插值法计算插值灰度对其进行尺度归一化处理,将图像尺度归一化为128×128,然后用直方图均衡化方法对图像进行灰度归一化处理,最后得到标准表情图像,为后续图像特征提取奠定基础。(2)提出一种融合局部二值模式和局部相位量的特征提取算法。该算法分别采用局部二值模式和局部相位量将人脸表情图像转换成相对应的特征谱,通过不同分块下的表情识别率实验,确定最优分块策略。接着将两种特征谱转换为直方图特征序列,并且将两个直方图特征序列串行连接,解决了局部特征提取表情特征信息丢失的问题。对比单独局部二值算法、局部相位量算法和二维Gabor小波算法,取得了较高的识别率。(3)提出一种基于深度信念网络模型的人脸特征提取与识别算法。对深度信念网络参数采用动态搜索方法,确定了最优的参数以及度信念网络深度与迭代次数。对比传统的支持向量机算法,该算法整体识别率提高了4.1%。(4)针对传统表情识别系统形式单一、可扩展性差、移动设备不兼容等问题。设计了一款基于B/S架构的在线人脸表情识别系统软件。解决了PC、平板、手机等设备兼容性问题,满足了响应式布局的设计理念,针对不同的设备视口,展现不同的页面样式。
[Abstract]:Facial expression is the carrier of human emotion, as a kind of important information in human daily communication, it reflects the psychological state of human beings. People transmit different emotional information to others through small expression changes. With the development of science and technology, The demands of human-computer interaction are more diversified. How to make the computer understand facial expression information has become a hot research topic. Nowadays, more and more attention has been paid to facial expression recognition technology. It is widely used in many fields. In this paper, the related algorithms used in the three steps of facial expression image recognition are studied in depth. The research contents and innovative work include the following points: 1) aiming at the preprocessing link. Cut pure expression image problem, Using integral projection direction to locate the precise area of human eye and eyeball. According to the traditional facial model of "three Chambers and five eyes", the pure expression image is cut out, and the bilinear interpolation method is used to calculate the interpolation gray level to normalize its scale. The image scale is normalized to 128 脳 128, and the histogram equalization method is used to normalize the gray level of the image. Finally, the standard facial expression image is obtained. A feature extraction algorithm combining local binary mode and local phase quantity is proposed, which uses local binary mode and local phase to transform facial expression image into corresponding feature spectrum. The optimal partitioning strategy is determined by the experiment of facial expression recognition rate under different blocks. Then, the two feature spectra are transformed into histogram feature sequences, and the two histogram feature sequences are connected serially. The problem of information loss of local feature extraction is solved. Compared with single local binary algorithm, local phase measure algorithm and two-dimensional Gabor wavelet algorithm, A face feature extraction and recognition algorithm based on depth belief network model is proposed. Dynamic search method is used for the parameters of depth belief network. The optimal parameters, the depth of degree and belief network and the number of iterations are determined. Compared with the traditional support vector machine algorithm, the overall recognition rate of the algorithm is improved by 4.1g / L) for the traditional expression recognition system, the form is single and the expansibility is poor. This paper designs an online facial expression recognition system software based on B / S architecture. It solves the compatibility problem of PC, tablet, mobile phone and other devices, satisfies the design idea of responsive layout, and aims at different device view ports. Display different page styles.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】