【作者】 李栋；

【导师】 田彦涛；

【作者基本信息】 吉林大学 ， 控制理论与控制工程， 2014， 博士

【摘要】 本文从表情识别的局部纹理信息分析入手，主要研究人脸关键点（如人眼）的定位、通过定位关键点（如人眼）来完成标准人脸图像的获取，进而将人脸面部表情放在局部特征的背景下进行识别，旨在达到两个目的：一是提高表情识别的准确率；二是提高算法的鲁棒性。此外，本论文还研究了一种新的基于决策树算法的人脸面部表情分类方法，提高了整个人脸面部表情识别的分类速度,为人脸面部表情识别分类研究提供了一种全新分析数据、预测数据的手段。同时，针对传统支持向量机分类参数采用网格寻优的方式，提出了一种利用粒子群算法对其进行自动优化的方法，提高了整个人脸面部表情识别系统分类的效率。本文主要针对静态人脸面部表情图像的以下方面进行了研究与探讨：如何定位人脸的关键部位、如何对人脸进行检测、人脸面部表情特征的提取与表征、如何对人脸面部表情进行分类。本研究的具体内容如下：1.提出了一种新颖高效的由粗到精的人眼检测策略，应用Gabor变换和PCA重构验证来粗略定位眼睛区域，再使用两级邻域运算对瞳孔进行精确定位。该方法具有非迭代和计算简单的优点，通过使用显著极值检测及PCA重构验证，省去训练样本的步骤，从而在保证定位精度的同时大大降低了计算量。实验结果表明，该方法的眼睛定位准确率可达到98.6%，实现了对眼睛的精确定位。2.采用Gabor滤波器提取人脸面部表情图像的多尺度和多方向特征,按照融合规则将相同尺度不同方向的Gabor模特征融合到一起。并利用LBP算子对Gabor模特征进一步编码，为了能够有效地表征图像全局特征,将采用子块分割法，即将融合图像分割为多个等面积且互相之间不重叠的小单元，对每一个单元内部的融合特征画出直方图，将所有的直方图结合起来分析表情图像。经实验验证，通过该方法提取的特征所得到的人脸面部表情识别率可达到93.42%，无论在计算量上还是识别性能上都比传统的特征提取更具有优势。3.针对传统支持向量机分类参数采用网格寻优的方式，提出了一种利用粒子群算法对其进行自动优化的方法，通过优化明确了最优的参数选择，提高了支持向量机的性能，经实验验证，该种优化的方法可以将识别率提高至96.05%，证明了此方法的有效性，提高了整个人脸面部表情识别系统分类的效率。此外，本文还将基于决策树算法运用到人脸面部表情分类中。不仅可以将其作为对比本文分类实验结果的一个依据而且还为人脸面部表情识别分类研究提供了一种全新分析数据、预测数据的手段。4.对全文所做的工作进行了总结，对文章中存在的不足以及今后需要研究问题的进行了说明。论文研究得到吉林省科技发展计划重点项目“基于混合特征的人脸表情识别方法与系统研究”（20071152），吉林大学“2010年研究生创新研究计划项目”（20101027）及吉林省青年科研基金项目“局部遮挡情况下的鲁棒表情识别方法研究”（20140520065JH）的资助。更多还原

【Abstract】 This paper studied the local texture feature of the expression recognition and the locationof key feature (such as eyes) was investigated on facial expression recognition in order torecognize the facial expressions based on the local feature. The purpose of this research is toimprove the facial recognition and the robustness of the algorithm. A novel facial expressionclassification based on the decision tree algorithm was provided in this paper, whichimproved the classification of facial expression recognition and provided an effectivemethod of data analysis and forecasts. The particle swarm optimization (PSO) algorithm wasused to solve the problem of traditional support vector machine (SVM) in order to improvethe accuracy of facial expression recognition system. For the static facial expression images,the location of the key facial feature, the face detection, the extraction and expression of thefeature of facial expression, and the classification were studied in this research. The mainresearch and innovative work are shown as follows:First, Gabor transform and principal component analysis (PCA) refactoring validationwere used to locate the eye area roughly and the two stage neighborhood arithmetic was usedto locate the pupil precisely. This method has the advantages of the non-iteratively and thesimple computation. Salient image extrema detection and PCA refactoring validation wereused to reduce the step of training sample computation under the requirement of locationaccuracy. The accuracy of eye location was more than98.6%as showed in the experimentresult.Second, Gabor filterbank was provided to extract the multi-scale and multi-directioncharacteristic of facial expression images. The features on the same scale in differentdirections of Gabor model syndrome were fused based on the fusion rules. The local binarypattern (LBP) operator was used to encode the Gabor model syndrome in. The sub-locksegmentation was used to improve the recognition of global image features. The image wassegmented into non-superposed elements with the equal area. Histograms of every element’sfusion features were combined to analyse the express image. The accuracy of facial expression recognition was93.42%as showed in the experiment. It has the advantage incomputation and recognition property compared with traditional method.Third, the PSO algorithm was provided in the traditional support vector machine withgrid search technique of classifying parameter to improve the efficiency of facial expressionrecognition system. The accuracy of recognition was improved to96.05%.Thedecision tree algorithm was used in classification of facial expression in this research whichprovided comparation of classification experiment and an effective method of data analysisand forecasts.Last, the main work of this research is summarized and the further work is discussed.This work was supported by Jilin Provincial Science and Technology development plankey Program (No.20071152), Graduate innovation project of Jilin University (No.20101027)and Jilin Youth Scientific Fund Project (No.20140520065JH).更多还原