人脸表情识别是一项极具发展潜力的生物特征识别技术。其研究目的是利用计算机进行人脸表情的识别，进而分析人的情感，能够进一步增强人机交互的友好性和智能性，因此人脸表情识别技术具有广阔的市场前景和应用价值。最近几年，人脸表情识别技术取得了前所未有的发展，但其在实际应用中的识别精度仍然难以满足人们的预期要求，主要是由于表情特征本身易受干扰、可区分性弱等原因，导致识别效果很难像人脸识别、指纹识别一样可靠。 本文研究了Gabor小波和局部二元模式(LBP)这两种人脸表情识别中常用的特征提取方法，利 用LBP算子对提 取的Gabor特征进行 编码，得到人 脸表情的局 部Gabor二值 模式(LGBP)。通过使用局部 保 全投影(LPP)算法，将提取得到的特征降维到LPP空间，使得识别过程简单而且只需要在低维空间计算，大大减少运算时间。经过对支持向量机和K近邻分类原理的研究发现，支持向量机的方法容易在样本的分界面附近出现错误分类，K近邻的分类方法则对分界面附近的样本可以实现很好的分类，但其在分类过程中计算量较大。根据以上研究，本文提出一种改进的方法，在分类界面附近实现对这两种方法融合，并通过实验在表情识别系统中进行验证。在同一表情库上进行实验可以看出，融合后的算法在识别率上要优于使用单一的支持向量机或者K近邻的分类方法。

Facial expression recognition is an extremely promising biometric technology. It is aiming at enhance the friendliness and intelligent of human-computer interaction by recognizing human facial expressions, and analyzing their emotions. Thus it has a wide range of application and possesses a practical value. In recent years, facial expression recognition technology has achieved unprecedented development, but its recognition accuracy in applications is still difficult to meet practical expectations, it is mainly due to vulnerability to interference and the weak discrimination of expression feature. So the performance is not as good as Face recognition and fingerprinting recognition. In this paper, we study two kinds of common facial expression recognition feature extraction algorithm, namely Gabor wavelet and Local Binary Pattern (LBP), and propose local Gabor binary Pattern (LGBP), it uses LBP to encode facial expression that extracted by Gabor features. And we project the extract characteristic to low-dimension by using local preservation projection (LPP) algorithm, it makes the algorithm more efficiency. By analyzing the theory of Support Vector Machine (SVM) and K nearest neighbor classifier, we find SVM has serious confusion in the vicinity of the sample interface, however, K nearest neighbor classifier can take advantage of information classification of samples near the surface, but the operation is rather high. This paper focuses on the method that combine them to improve the performance of both algorithms. And the result of simulation indicates the superior performance of the proposed approach.