基于多尺度中心化二值模式的人脸表情识别

现有局部二值模式(LBP) 算子存在不足: 产生的直方图维数过长、鉴别力不高、对噪声反应敏感. 针对此类问题, 提出中心化二值模式(CBP) 算子, 其优点: 1) 通过比较邻域中近邻点对, 大大降低了直方图维数; 2) 考虑中心像素点的作用并赋予其最高权重, 实现鉴别力的提高; 3) 改变LBP算子的符号函数, 明显减弱白噪声对图像的影响.此外, 为提高识别率, 将多尺度CBP(MCBP) 直方图作为人脸表征. 为增强算法对表情图像中细小变形的鲁棒性, 引入图像欧式距离(IMED) 并将其嵌入MCBP方法. 在JAFFE和Cohn-Kanade表情库的实验结果表明: 所提方法优于其它表情识别方法, IMED可增强MCBP的表情识别能力.

Facial expression recognition based on multi-scale centralized binary pattern

The existing local binary pattern(LBP) operators have disadvantages of long histograms produced by them, low discrimination and high sensitivities to noise. To deal with these problems, we propose the centralized binary pattern(CBP) operator. The CBP operator has several advantages: 1) It significantly reduces the histogram dimensionality by comparing pairs of neighbors in the neighborhood; 2) It enhances the discrimination by emphasizing the effect of the center pixel point through giving it the largest weight; 3) It decreases the white noise influence on face images by modifying the sign function of the existing LBP operator. Moreover, the multi-scale CBP(MCBP) histogram is used as face representation to increase the recognition accuracy. Furthermore, in order to improve the robustness to small deformation of expressional images, the image Euclidean distance(IMED) is introduced and embedded in MCBP. Experiments on JAFFE and Cohn-Kanade facial expression databases demonstrate that the proposed method outperforms other modern approaches and show that IMED can enhance the performance of MCBP in facial expression recognition.