【摘要】：人臉面部表情識別是通過計算機對人臉面部由肌肉拉動所產(chǎn)生的表情圖像或視頻做特征提取工作,并按照人類目前理解經(jīng)驗和思想認識來實施表情歸類和表情識別,從面部信息中提取分析人類情感。表情特征提取的正確性和有用性是表情可否正確識別的關鍵。本論文的重點是對表情圖像特征提取方法進行研究。本論文主要工作具體有以下幾個方面:首先,詳細介紹人臉表情識別系統(tǒng)各功能模塊,研究了圖像獲取模塊和預處理模塊的原理與算法,并進行小樣本采集實驗,包括以下四個方面:人臉檢測、圖像灰度化、圖像歸一化、光照補償。其次,對比研究三種常見的表情特征提取算法,包括:局部二值模式(Local Binary Pattern,LBP)、局部相位量化(Local Phase Quantization,LPQ)、旋轉不變局部相位量化(Rotation Invariant Local Phase Quantization,RILPQ),在JAFFE圖像庫部分圖像特征提取實驗,提取到特征向量灰度圖及量化直方圖做研究比對。本論文在RILPQ算法基礎上,引入二維高斯核方向導數(shù),提出一種新的特征提取算法,即:融合高斯導數(shù)RILPQ算法。再次,研究支持向量機(upport Vector Machine,SVM)理論并運用SVM模式識別與回歸的軟件包LIBSVM完成分類識別與回歸。本部分主要研究基于同向高斯核方向導數(shù)與RILPQ融合的人臉表情特征提取算法程序設計,并對三個參數(shù)做大量實驗研究,包括:方向導數(shù)濾波方向、濾波尺度、尺度半徑,尋找到一組最佳實驗參數(shù),表情識別率最高為92.57%。同時,為驗證該算法實驗效果,通過運行時間和表情識別率兩項指標與前面的三種特征提取算法進行比較,證明該算法運行時間較長但是能取得較好的表情識別分類效果。最后,本文又提出了一種異向高斯核方向導數(shù)與RILPQ融合的運動模糊人臉表情特征提取算法。通過JAFFE圖像庫水平方向運動模糊處理后做特征提取進行表情分類。實驗證明:在模糊長度為5像素,尺度半徑為R=9條件下,運動模糊表情識別率為66.10%,優(yōu)于RILPQ算法識別率1.4個百分點。
[Abstract]:Facial expression recognition is to extract facial expression images or videos generated by muscle pull by computer, and to classify and recognize facial expressions according to the current understanding experience and ideological understanding of human beings. Extract and analyze human emotion from facial information. The correctness and usefulness of expression feature extraction is the key to the correct recognition of expression. The focus of this paper is to study the method of facial expression image feature extraction. The main work of this paper is as follows: firstly, the functional modules of facial expression recognition system are introduced in detail, the principle and algorithm of image acquisition module and preprocessing module are studied, and the small sample acquisition experiment is carried out. Including the following four aspects: face detection, image grayscale, image normalization, lighting compensation. Secondly, three common expression feature extraction algorithms are compared, including: local binary mode (Local Binary Pattern,LBP), local phase quantification (Local Phase Quantization,LPQ), rotation invariant local phase quantification (Rotation Invariant Local Phase Quantization,RILPQ). In the experiment of image feature extraction in JAFFE image database, the feature vector grayscale map and quantitative histogram are extracted and compared. In this paper, based on the RILPQ algorithm, the directional derivative of two-dimensional Gaussian kernel is introduced, and a new feature extraction algorithm is proposed, that is, the fusion of Gao Si derivative RILPQ algorithm. Thirdly, the theory of support vector machine (upport Vector Machine,SVM) is studied and the classification recognition and regression are completed by using the software package LIBSVM of SVM pattern recognition and regression. In this part, we mainly study the program design of facial expression feature extraction algorithm based on the fusion of Gao Si kernel guide number and RILPQ, and do a lot of experimental research on three parameters, including directional derivative filtering direction, filtering scale, scale radius. A set of optimal experimental parameters were found, and the highest expression recognition rate was 92.57%. At the same time, in order to verify the experimental effect of the algorithm, the running time and expression recognition rate are compared with the previous three feature extraction algorithms, and it is proved that the algorithm has a long running time but can achieve better expression recognition and classification effect. Finally, this paper proposes a motion fuzzy facial expression feature extraction algorithm based on the fusion of abnormal Gao Si kernel guide number and RILPQ. The expression classification is carried out by feature extraction after horizontal motion blur processing in JAFFE image database. The experimental results show that under the condition that the fuzzy length is 5 pixels and the scale radius is R 鈮，

本文編號：2494953