【摘要】 针对人脸的静态图片不能描述表情动态信息的局限性,提出了一种基于注意力机制的视频人脸表情识别算法。首先,为了减少图像特征在处理过程中的损失,在VGGNet16模型的侧方添加一系列卷积核,形成一个双向监督模块,同时利用上采样、下采样对各个侧输出层的特征图进行加权融合改进;其次,利用改进后的VGGNet16模型提取视频序列图片的空间特征与使用光流法提取图片中的时间特征进行融合;然后,使用注意力机制对其进行特征加权,在LSTM网络中对加权后的特征进行训练和分类;最终,该模型在AFEW数据集和CK+数据集上的识别率分别为6111%和958%,与现有文献识别率的对比验证了文中算法的优势。更多还原

【Abstract】 Aiming at the limitation that the static image of the face can not describe the limitation of the dynamic information of the expression,a video facial expression recognition algorithm is proposed based on the attention mechanism. Firstly,in order to reduce the loss of image features in the process,a series of convolution kernels are added on the side of the VGGNet16 model to form a two-way supervised module.At the same time,the up-sampling and down sampling are used to improve the weighted fusion of the feature maps of each side output layer. The improved VGGNet16 model is used to extract the spatial features of the video sequence image,and the optical flow method is used to extract the temporal features in the image. The attention mechanism is used to weight the features and train the weighted features in the LSTM network and classification. Finally,the recognition rates of the model on the AFEW dataset and CK+ dataset are 61.11% and 95.8%,respectively. The comparison with the existing literature recognition rate verifies the advantages of the proposed algorithm.更多还原