基于多尺度融合特征网络的表情识别研究

Expression Recognition Based on Multi-Scale Fusion Feature Network

在人脸表情识别任务中场景和表情数据丰富且复杂的情况下,卷积神经网络难以提取具有代表性的表情特征,因此提出一种多尺度融合特征网络。首先,在卷积神经网络前端引入具有不同大小卷积块的Inception V2结构,既增强了网络模型提取表情图片局部特征的能力,又减少了网络模型的训练参数量;然后,利用Grad-CAM热力权重可视化技术绘制热力权重分布图,通过1×1卷积块构建融合特征,使其同时兼具浅层局部特征和深层语义特征;最后,将BN结构和Dropout结构引入改进网络结构中,以防止模型出现过拟合或欠拟合问题。在公开数据集FER2013和融合数据集(CK+、JAFFE和RaFD)上进行实验,结果表明该方法的识别准确率更高、泛化能力更强。

A multi-scale fusion feature network is proposed for the problem that convolutional neural network is difficult to extract representative expression features in the case of rich and complex scene and expression data in face expression recognition tasks.First,the Inception V2 structure with different size convolutional blocks is introduced in the front end of the convolutional neural network,which enhances the ability of the network model to extract local features of expression pictures and reduces the number of training parameters of the network model.Then,the Grad-CAM thermal weight visualization technique is used to draw the thermal weight distribution,and the fusion features are constructed by 1×1 convolutional blocks,so that they include both shallow local features and deep semantic features at the same time.Finally,the BN structure and Dropout structure are introduced into the improved network structure to prevent the model from overfitting or underfitting problems.Experiments are conducted on the public dataset FER2013 and the fused datasets(CK+,JAFFE and RaFD),and the results show that the method has higher recognition accuracy and better generalization ability.