FSAGN:一种自主选择关键帧的表情识别方法  

FSAGN:An expression recognition method based on independent selection of video key frames

作　　者：祝锦泰 叶继华[1] 郭凤 江蕗 江爱文[1] ZHU Jintai;YE Jihua;GUO Feng;JIANG Lu;JIANG Aiwen(School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330022,China;Department of Information Engineering,Zibo Technician College,Zibo 255030,Shandong Province,China)

机构地区：[1]江西师范大学计算机信息工程学院,江西南昌330022 [2]淄博技师学院信息工程系,山东淄博255030

出　　处：《浙江大学学报（理学版）》2022年第2期141-150,158,共11页Journal of Zhejiang University（Science Edition）

基　　金：国家自然科学基金资助项目(61462042,61966018).

摘　　要：由于在包含表情的视频数据集中存在大量与表情特征无关的视频帧,使得模型在训练中学习到大量无关信息,导致识别率大幅下降,因此如何令模型自主地选择视频关键帧成为研究的关键。在已有的视频表情识别方法中,大多没有考虑关键帧和非关键帧对模型训练效果的影响,为此提出了一种基于注意力机制与GhostNet的人脸表情识别(FSAGN)模型。通过自注意力机制与帧选择损失计算不同帧的权重,根据权重自主选择视频序列的关键帧。此外,为减少模型参数、降低模型的训练成本,将传统的特征提取网络替换为训练参数较少的GhostNet网络,并与注意力机制结合,分别在CK+和AFEW数据集中进行了实验,得到的最高识别率分别为99.64%和52.31%,分类正确率具有竞争力,适用于对视频序列较长且在视频序列中表情特征分布不均匀的面部表情识别。As there exist a large number of video frames unrelated to facial expressions in the video data set containing facial expressions,a large amount of information unrelated to facial expressions is learned in the training process of the model,which results in a significant decline of the performance.So how to make the model capable of choosing the relevant video key frame autonomously becomes the key problem.At present,most of the existing video expression recognition methods do not yet consider the different effects of key frame and non-key frame on the training effect of the model.In the paper,a face expression recognition model based on attention mechanism and GhostNet(FSAGN)is proposed.The model calculates the weights of different frames by self-attention mechanism and frame selection loss,then selects the key frames of the video sequence autonomously according to the weights.In addition,in order to reduce model parameters and training costs,our approach replaces the traditional feature extraction network with the GhostNet network with fewer training parameters,and combines it with the attention model.Experiments were carried out on the designed network in CK+and AFEW data sets,and the highest recognition rates were 99.64%and 52.31%,respectively,which reached a competitive classification accuracy.It was suitable for facial expression recognition tasks with long video sequences and uneven distribution of facial expression features in video sequences.

关 键 词：面部表情识别 注意力机制 关键帧自主选择 GhostNet 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]