Cai Weiwei, Gao Ming, Liu Runmin, Mao Jie
College of Sports Engineering and Information Technology, Wuhan Sports University, Wuhan, China.
School of Logistics and Transportation, Central South University of Forestry and Technology, Changsha, China.
Front Psychol. 2021 Oct 22;12:762795. doi: 10.3389/fpsyg.2021.762795. eCollection 2021.
Understanding human emotions and psychology is a critical step toward realizing artificial intelligence, and correct recognition of facial expressions is essential for judging emotions. However, the differences caused by changes in facial expression are very subtle, and different expression features are less distinguishable, making it difficult for computers to recognize human facial emotions accurately. Therefore, this paper proposes a novel multi-layer interactive feature fusion network model with angular distance loss. To begin, a multi-layer and multi-scale module is designed to extract global and local features of facial emotions in order to capture part of the feature relationships between different scales, thereby improving the model's ability to discriminate subtle features of facial emotions. Second, a hierarchical interactive feature fusion module is designed to address the issue of loss of useful feature information caused by layer-by-layer convolution and pooling of convolutional neural networks. In addition, the attention mechanism is also used between convolutional layers at different levels. Improve the neural network's discriminative ability by increasing the saliency of information about different features on the layers and suppressing irrelevant information. Finally, we use the angular distance loss function to improve the proposed model's inter-class feature separation and intra-class feature clustering capabilities, addressing the issues of large intra-class differences and high inter-class similarity in facial emotion recognition. We conducted comparison and ablation experiments on the FER2013 dataset. The results illustrate that the performance of the proposed MIFAD-Net is 1.02-4.53% better than the compared methods, and it has strong competitiveness.
理解人类情感和心理是实现人工智能的关键一步，而正确识别面部表情对于判断情感至关重要。然而，面部表情变化所引起的差异非常细微，不同的表情特征较难区分，这使得计算机难以准确识别人类面部情感。因此，本文提出了一种具有角距离损失的新型多层交互式特征融合网络模型。首先，设计了一个多层多尺度模块来提取面部情感的全局和局部特征，以便捕捉不同尺度之间的部分特征关系，从而提高模型区分面部情感细微特征的能力。其次，设计了一个分层交互式特征融合模块，以解决卷积神经网络逐层卷积和池化导致有用特征信息丢失的问题。此外，还在不同层次的卷积层之间使用了注意力机制。通过增加各层上不同特征信息的显著性并抑制无关信息，提高神经网络的判别能力。最后，我们使用角距离损失函数来提高所提出模型的类间特征分离和类内特征聚类能力，解决面部情感识别中类内差异大、类间相似度高的问题。我们在FER2013数据集上进行了比较和消融实验。结果表明，所提出的MIFAD-Net的性能比比较方法好1.02 - 4.53%，具有很强的竞争力。
