采用支路辅助学习的人脸表情识别

人脸表情识别属于一种细粒度识别，模型需要同时聚焦于浅层与深层特征。针对独立结构的卷积神经网络对细粒度特征的提取、融合能力不足的问题，提出一种基于支路辅助学习的网络结构。在基础网络的输入层引入一条支路辅助网络，该网络将逐层使用金字塔卷积块提取全局特征；通过特征映射模块不断将支路提取到的决策信息传导至基础网络，辅助基础网络提取细节特征；在模型输出层采用特征拼接的方式将支路网络与主路网络融合。将所提出的网络在公开人脸表情数据集CK+、JAFFE、FER2013和MMEW上进行识别实验，结果表明：支路辅助学习模块能够有效提升基础网络的特征提取能力和泛化能力，提出的方法识别率达到了98.89%、94.80%、71.88%和86.67%，比仅采用基础网络（例如：ResNet50）进行识别提高了3.49、2.2、5.51和1.48个百分点。

Facial Expression Recognition Method Based on Branch-Assisted Learning Network

Facial expression recognition is a kind of fine-grained recognition, this model needs to focus on both shallow and deep features. Aiming at the problem that the convolutional neural network with independent structure has insufficient ability to extract and fuse fine-grained features, a network structure based on branch-assisted learning is proposed in this paper. Firstly, a branch network is introduced into the input layer of the basic network, which will extract global features by using pyramid convolution blocks. Then, through the feature-mapping module, the decision information extracted by the branch is continuously transmitted to the basic network to assist the basic network to extract the detailed features. Finally, in the output layer of the model, the branch network and base network are fused by feature stitching. The recognition experiments of the proposed network on the public facial expression data sets CK+, JAFFE, FER2013 and MMEW show that the branch-assisted learning module can effectively improve the feature extraction ability and generalization ability of the basic network, and the recognition rate of the proposed method reaches 98.89%, 94.80%, 71.88% and 86.67%, which is 3.49, 2.2, 5.51 and 1.48 percentage points higher than that of only using the basic network（e.g.ResNet50）.