摘要
针对传统局部二元模式(LBP)算子存在直方图维数过高而导致识别速度降低和二值数据对噪声很敏感的问题,在分析传统LBP算子的原理基础上,对人脸表情特征的数据量增加、人脸表情特征向量和特征识别过程的优化进行如下改进:将人脸表情图像经过小波包的分解和重构,得到4幅不同频段的图像,从而有效地增加原表情图像的数据量;采用修正的LBP算法对人脸表情图像进行特征提取,并通过稀疏表示模型优化其特征向量,有效地降低传统LBP直方图的维数,提高人脸表情识别率,二次修正的LBP算法鲁棒性好;构建基于神经网络的多分类器模型,融合多特征多分类器的输出,有效地提高表情特征分类的准确性和稳定性。研究结果表明:与传统LBP算法对比,本算法用于人脸表情的识别时,其识别率得到较大幅度提高,算法鲁棒性好。
Based on the low discrimination rate induced by the high dimension of image histogram and the binary data sensitivity to noise on the traditional local binary pattern （LBP） operators, the optimization of rich data, construction and recognition feature vector for the facial expression feature were improved after analyzing the principle of LBP operator. The original facial expression image was decomposed and reconstructed with wavelet packet to get 4 images at different frequency bands which effectively increase the data amount of original image. The LBP algorithm was adapted to extract features from face images and its eigenvector was optimized by using the sparse representation model, which effectively reduces the dimension of the traditional LBP histogram and improves the facial expression recognition, and as a result, the algorithm is robust. The multi-classifier was modeled based on the neural network model and decision fusion was implemented for the output of the multiple features classifier, which improves the accuracy and stability of expression feature classification. The results show that this algorithm is suitable for facial expression recognition compared with the traditional LBP algorithm and the recognition rate is improved significantly, and the algorithm is robust.
出处
《中南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2014年第5期1503-1509,共7页
Journal of Central South University:Science and Technology
基金
湖南省自然科学基金资助项目(12GJ6055)
关键词
LBP算子
自动识别
鲁棒性
纹理
LBP operator
automatic identification
robustness
texture
