改进LDP结合几何特征融合的人脸表情识别  

Facial expression recognition method based on improved LDP and geometric feature fusion

作　　者：成亚丽[1] 秦飞龙 李政文 CHENG Ya-li;QIN Fei-long;LI Zheng-wen(School of Big Data and Artificial Intelligence,Chengdu Technological University,Chengdu 611730,China;Department of Arts and Sciences,Chengdu College of University of Electronic Science and Technology of China,Chengdu 611730,China)

机构地区：[1]成都工业学院大数据与人工智能学院,四川成都611730 [2]电子科技大学成都学院文理系,四川成都611730

出　　处：《计算机工程与设计》2021年第9期2577-2584,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61533006);四川省科技厅重点攻关基金项目(2019YJ0375)。

摘　　要：通过局部方向模式(LDP)提取的局部外观特征常用于人脸表情识别中,但其存在容易受噪声像素影响和伪编码的问题。为此,提出一种改进型的LDP编码方案(ILDP)。利用Sobel算子代替Kirsch掩模来提取图像中的梯度信息,将梯度信息进行对数变换后进行累加,避免噪声像素点的影响;通过采用一个梯度幅度阈值判断是否为平坦区域,避免进行错误的LDP编码。为进一步提高识别率,将ILDP提取的局部外观特征与主动表观模型(AAM)提取的全局几何特征相结合,通过PCA进行降维。在JAFFE和BU-3DFE两个公共数据集上的实验结果表明,该方法能够有效提高表情的识别率,降低对噪声的敏感性。The local appearance feature extracted using LDP is often used in facial expression recognition,but it is easy to be affected by noise pixels and pseudo coding.Therefore,an improved LDP coding scheme(ILDP)was proposed.Sobel operator was used to extract the gradient information in the image instead of Kirsch mask,and the gradient information was logarithmically transformed and accumulated to avoid the influence of noise pixels.A gradient amplitude threshold was used to determine whe-ther it was a flat region,avoiding incorrect LDP coding.To further improve the recognition rate,the local appearance feature extracted using ILDP was combined with the global geometry feature extracted using AAM,and the dimension was reduced using PCA.Experimental results on databases JAFFE and BU-3DFE show that,the proposed method can effectively improve the recognition rate of expression and reduce the sensitivity to noise.

关 键 词：面部表情识别 局部方向模式 噪声像素 伪编码 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]