对噪声鲁棒性的表情特征提取方法  

Robust expression feature extraction method for noise

作　　者：韩肖 马祥[1] HAN Xiao;MA Xiang(School of Information Engineering,Chang’an University,Xi’an 710064,China)

机构地区：[1]长安大学信息工程学院,陕西西安710064

出　　处：《电子设计工程》2022年第13期1-5,共5页Electronic Design Engineering

基　　金：国家自然科学基金(61771075);中央高校基本科研业务费资助项目(300102249203)。

摘　　要：为了有效解决在噪声污染下人脸表情识别的问题,提出了一种具有鲁棒性的表情特征提取方法。通过使用不同的低秩投影矩阵和映射矩阵的方式来约束低秩稀疏表示的重构图像,将高维空间中的原始数据映射到低维空间模型中,以提取出更具判别性的鲁棒特征。同时所提出的方法采用L_(2,1)范数进行约束,L_(2,1)范数不仅对含噪数据和离群数据具有鲁棒性,还具有旋转不变性,能够较好地保留数据的几何结构,这在一定程度上可以提高识别性能。CK+、JAFFE及LFW数据集的实验结果证明了该文方法对噪声具有鲁棒性。In order to effectively solve the problem of facial expression recognition under noise pollution,a robust facial expression feature extraction method is proposed.By using different low-rank projection matrices and mapping matrices to constrain the low-rank sparse representation of the reconstructed image,the original data in the high-dimensional space is mapped to the low-dimensional space model to extract more discriminative robustness feature.And the proposed method uses the L_(2,1)norm to constrain.The L_(2,1) norm is not only robust to noisy data and outlier data,but also has rotation invariance and can well retain the geometric structure of the data.To a certain extent,the recognition performance can be improved.The experimental results of CK+,JAFFE and LFW datasets prove that the proposed method is robust to noise.

关 键 词：表情识别 特征提取 低秩稀疏表示 低维映射 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]