【作者】 余锐；

【导师】 王茜；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 在整个人类社会交际过程中,人脸表情充当着一种重要的信息传递媒介的角色,对协调人际关系有着重要的作用。心理学家A.Mehrabian通过研究发现,在人类日常生活中,人脸表情所传达的信息比重占据了总信息数的一半以上。因此,人脸表情的研究也是比较热门,自上世纪70年代,自动人脸表情特征分析驱动着大量的研究,其中包括人机交互、智能控制、机器视觉、情感计算等方面,其中主要集中在表情自动识别这一方向。因为通过对表情特征的分类识别能够取得对科研或者商业上很重要的价值,而后可以重新利用获取到的人脸表情信息来发掘更大的包括商业以及其他方面更大的价值。本文参考大量国内外文献资料,基于自动人脸表情特征分析相关研究背景、意义以及相关技术进行了详尽的分析与论述,前期对传统的图像识别技术所运用到的技术进行了分析,其中主要集中在图像处理,体现在对最开始样本进行前期处理工作,其中主要目的是减少光照以及相关噪声对后期所带来的实验数据采集中的误差,可以采取灰度预处理以及几何预处理,如旋转裁剪、图像定位、几何光照归一化等,进行前期的处理后更利于后面具体表情的特征提取和分类准确率的提高,随后就是针对前期处理好的数据集进一步的进行特征提取和选择,包括之前相关研究所采用的Gabor小波变换、SIFT以及PCA降维等进行表情图像特征的提取,然后可以通过Adaboost算法进行特征选择,最后采取比较流行的分类器算法其中包括KNN、SVM、神经网络以及HMM进行表情特征的分类。之后采取了基于深度学习的CNN网络模型,应用到表情数据集的识别中,之前提出本文方法基于典型特征区域以及LBP纹理算子的特征提取,并根据实验结果来进行网络模型参数与结构的优化,从而达到一个最优化模型标准,并利用深度学习包括模型泛化能力强,实验操作简单等优势,与之前的图像识别方法进行了对比,以此体现出基于深度学习CNN表情图像识别的特点,主要集中在高的识别准确率以及易操作性。整个实验在Tensorflow平台下进行,实现多分类人脸表情的识别。更多还原

【Abstract】 In the whole social communication process of human society,facial expression plays an important role in information transmission media,and plays an important role in coordinating interpersonal relationship.Psychologist A.Mehrabian finds it through research,in human daily life,the proportion of information conveyed by facial expression occupies more than half of the total information.Therefore,the research on facial expression is also popular,since the last century in 70 s,automatic facial expression analysis of the characteristics has driven a lot of researches,including human-computer interaction,intelligent control,machine vision,affective computing,etc.,it mainly focuses on automatic expression recognition.Because the classification and recognition of facial expression features can be very valuable for scientific research or business,and then we can re-use the obtained facial expression information to discover greater value,including business and other aspects.This paper references to the data of a large number of domestic and foreign literature,based on automatic facial expression feature analysis,the relevant research background,significance and related technologies are analyzed and discussed in detail.In the early stage,the technology applied in traditional image recognition technology was analyzed,which mainly focus on image processing,it is reflected in the preliminary processing of the initial samples.,the main purpose is to reduce the data acquisition of light and noise brought by the late in error,grey preprocessing and geometric preprocessing can be adopted,Such as rotary cutting,image location,geometric lighting normalization and so on,early treatment is more conducive to feature extraction and classification accuracy of the following expressions is improved,then for the pre-processed data set for further feature extraction and selection,including the extraction before the relevant research institute of the Gabor,wavelet transform SIFT and PCA dimension of facial expression image features,and then through the Adaboost algorithm for feature selection,and finally take the popular points Class algorithms include KNN,SVM,neural networks,and HMM for classification of facial features.After taking a CNN network model based on deep learning,applied to the face recognition data sets,the method of feature extraction based on typical feature regions and LBP texture operator is proposed before,and to optimize the parameters and structure of network model based on the experimental results,so as to achieve a standard optimization model,and the use of deep learning including model generalization ability,the simple operation and other advantages,compared with the previous image recognition method,in order to reflect the characteristics of CNN face image recognition based on deep learning,mainly in the high accuracy and easy operation.The whole experiment is carried out under the Tensorflow platform to realize multi classification facial expression recognition.更多还原