基于二阶有效通道注意力网络的无约束人脸表情识别

现有基于卷积神经网络的无约束人脸表情识别方法侧重于网络结构的设计，对网络学习到的通道间深层特征相关性的研究较少，没有充分利用神经网络提取表达力较强的特征。为解决此问题，设计了一种基于二阶有效通道注意力网络（second-order efficient channel attention network，SECA-Net）的无约束人脸表情识别方法。该方法采用轻量级的网络提取表情图像的深层特征，使用二阶有效通道注意力模块统计深层特征的二阶信息并捕捉跨通道特征间的依赖关系来自适应地缩放通道特征，进而获得更具判别力的表情特征。SECA-Net利用Softmax损失和中心损失联合优化模型进行表情分类，该模块具有较少的参数量、较低的显存需求和计算量，并且没有使用额外的数据预训练模型。同时，所提出的模块还能提取到人脸表情微小变化的局部特征。在RAF-DB和FER-2013无约束人脸表情数据集上的实验结果表明，提出的方法是有效的。

Second-order efficient channel attention network for unconstrained facial expression recognition

Existing convolutional neural networks (CNN)-based unconstrained FER methods focus more on the design of the network architecture, and ignore the research on the deep feature correlations between different channels, which prevents the neural network from extracting the features with strong expressive ability. In order to address the above issue, a second-order efficient channel attention network (SECA-Net) is proposed. SECA-Net consists of a lightweight feature extraction network and a SECA module. Firstly, the lightweight network is used to extract deep expression features. Then, SECA module is designed to achieve second-order statistics of deep features and adaptively rescale the channel-wise features by capturing the dependencies of cross-channel features for more discriminative feature representation. Finally, Softmax loss and central loss are jointly used to optimize the model for expression classification. SECA-Net has lightweight parameters, low memory requirements and calculations, and does not use external database to pre-train the model. Moreover, the proposed SECA module can extract the local features of small facial expression changes. The experimental results on two unconstrained datasets (RAF-DB and FER-2013) demonstrate the effectiveness of the proposed method.