基于多特征与卷积神经网络的人脸表情识别

摘要

A novel facial expression recognition method based on multiple feature extraction and convolution neural network is proposed.In the preprocessing phase,expression images are cropped by using "three court five eyes" and the geometric model of the face,scaled by using bicubic interpolation.Three different features are extracted:local direction patterns,discrete wavelet transform and sobel operator.These three feature images are input into the convolution neural network in the form of three channel images for adaptive fusion.And the features are classified by the Softmax layer.The recognition rate of this method in CK + is 99.5％ and in RAF-DB is 72.1％,the recognition rate of the two databases is improved.The results show the effectiveness of the method proposed.%提出一种多特征与卷积神经网络相结合的人脸表情识别方法.先对人脸表情图像进行预处理,根据人脸面部“三庭五眼”的特征和人脸的几何模型对图像进行裁剪,采用双三次插值法对图像进行缩放.然后提取样本的局部方向模式、二维离散小波变换、Sobel算子三种特征.将这三种特征以三通道图像的形式输入卷积神经网络中进行自适应融合,融合后的特征通过Softmax层进行分类.在CK+数据库的识别率为99.51％,在RAF-DB的识别率为72.1％,识别率都有所提升,验证了所提方法的有效性.