Abstract: A multi-scale integrated attention network（MIANet） is proposed to address the problems of difficulty in extracting effective features and complex network model parameters in the current ordinary convolutional neural network for facial expression recognition. Firstly, in order to increase the width and depth of the network while avoiding redundant calculations, an Inception structure is introduced into the network, which can be used to extract multi-scale feature information of images. Then, the efficient channel attention（ECA） mechanism emphasizes the regions associated with facial expression and suppresses the irrelevant background regions to improve the representation ability of important facial features. Finally, deep separable convolution is used in the convolution layer to reduce network parameters and prevent over fitting. Experiments on public data sets FER-2013 and CK+ with the proposed method show 95.76% and 72.28% accuracy, respectively. The experimental results show that the method has a good recognition effect and strong generalization ability, and it has a certain reference value in facial expression recognition in terms of network structure setting and parameter configuration.

Key words: facial expression recognition, multi-scale feature extraction, deep separable convolution, attention mechanism

摘要： 针对在人脸表情识别中普通卷积神经网络难以提取有效特征、网络模型参数复杂等问题，提出了一种多尺度融合注意力机制网络（multi-scale integrated attention network，MIANet）。为了同时增加网络的宽度和深度又避免冗余计算，在网络中引入Inception结构，用于提取图像的多尺度特征信息。使用高效通道注意机制（efficient channel attention，ECA），强调与面部表情相关的区域抑制不相关的背景区域，提高重要面部特征的表达能力。在卷积层中采用深度可分离卷积，减少网络参数，防止过拟合。使用提出的方法在公开数据集FER-2013和CK+上进行实验，分别取得了95.76%和72.28%的准确率。实验结果表明，该方法识别效果较好，泛化能力较强，在人脸表情识别中对网络结构设置和参数配置方面具有一定的参考价值。

关键词: 人脸表情识别, 多尺度特征提取, 深度可分离卷积, 注意力机制