基于多视角深度网络增强森林的表情识别  

Multi-view Deep Neural Network Enhanced Forests for Facial Expression Recognition

作　　者：张发勇[1] 刘袁缘 李杏梅[2] 覃杰 Zhang Fayong;Liu Yuanyuan;Li Xingmei;Qin Jie(Faculty of Information Engineering,China University of Geosciences,Wuhan 430074;School of Mechanical Engineering and Electronic Information,China University of Geosciences,Wuhan 430074;Wuhan Huazhong Numerical Control Co,Ltd,Wuhan 430074)

机构地区：[1]中国地质大学信息工程学院,武汉430074 [2]中国地质大学机械与电子信息学院,武汉430074 [3]华中数控股份有限公司红外事业部,武汉430074

出　　处：《计算机辅助设计与图形学学报》2018年第12期2318-2326,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61602429)

摘　　要：为了提高在自然环境中姿态变化下人脸表情识别的准确性和鲁棒性,提出一种基于多视角深度网络增强森林的表情识别方法.首先提取人脸区域的人脸子块以消除人脸遮挡等噪声影响,通过在预训练的卷积神经网络模型上迁移学习获得深度表情特征;然后,估计水平自由度下的头部姿态参数以消除头部姿态运动的影响,建立多视角条件概率模型,并将条件概率和神经联结函数引入随机树的节点分裂学习中,提高模型在有限训练集上的学习能力和区分力;最后通过多视角权重投票决策人脸表情类别.M-DNF能够获得不同视角下的表情分类结果,而不需要大量的数据集训练.在CK+、多视角BU-3DFE和自发LFW这3个具有挑战的公共人脸数据集上进行实验的结果表明,该方法平均识别准确率分别达到98.85%, 86.63%和57.20%,均高于目前已有且公认的识别率高的表情识别方法.In order to improve the accuracy of multi-view facial expression recognition in natural environment,a novel multi-view deep neural network enhanced random forest(M-DNF)is proposed for robust facial recognition.First,our method extracts robust deep transfer expression features from random facial patches to reduce the influence from various noises,such as occlusion,etc.Then,in order to eliminate the influence of pose various,the M-DNF is devised to enhance decision trees with the capability of representation learning from transferred convolutional neural networks and to model facial expression of different views with conditional probabilistic learning.M-DNF can achieve both head poses and facial expressions,and performs well even when there are only a small amount of training data.Experiments were conducted using public CK+,multi-view BU-3DFE and LFW datasets.Compared to the state-of-the-art methods,the proposed method achieved much improved performance and great robustness with an average accuracy of 98.85%on CK+facial datasets,86.63%on the multi-view BU-3DEF dataset,and 57.20%on LFW in-the-wild dataset.

关 键 词：人脸表情识别 多视角深度网络增强森林 头部姿态配准 深度迁移特征学习 

分 类 号：TN911.73[电子电信—通信与信息系统]