基于四元数局部编码和卷积网络的表情识别

Expression recognition based on quaternion local coding and convolutional network

为充分利用彩色图像的颜色信息和通道之间的关联性,提出一种联合四元数矩阵相位信息和幅值信息的特征提取方法,结合卷积神经网络(CNN)进行表情识别。将彩色表情图像表示为纯四元数矩阵并进行Clifford平移,对相位和幅值分别进行局部二值模式(LBP)编码,提取多尺度融合的图像特征,将特征输入CNN进行训练并分类。实验结果表明,该算法在RafD和MMI表情库上的识别率分别为79.42%和93.28%,相比其它表情识别算法,识别率更高且识别效果稳定。

To make full use of the color information of color image and the correlation among different channels,an approach of feature extraction combining the amplitude and phase information of quaternion matrix was proposed,which was used for expression recognition together with convolutional neural network(CNN).Color facial images were represented into pure quaternion matrices and operated with Clifford translation.The weighted phase and amplitude information were respectively encoded using LBP to extract the multiscale fused features.The fusion features were served as the input of CNN for training and classification.Results of experiments performed on RafD and MMI facial expression databases show that the recognition rates are 79.42%and 93.28%respectively.Compared with other facial expression recognition algorithms,the proposed algorithm has higher recognition rate and its recognition performance is stable.