Aiming for the ResNet network structure used for feature extraction, there are still problems of overfitting, slow calculation speed and room for improvement of accuracy. A mixed depthwise separation residual network is proposed for image feature extraction. The residual and inverted residual blocks are used in the research, which makes the three kinds of residual units iteratively interact in the training process. This process promotes the feature representation of deep learning neural networks and alleviates the problem of feature dispersion caused by the deep network level in traditional expression recognition research. The experiment used CK+ and fer2013 datasets. The synthesis of multiple experimental results shows that the optimized deep learning model has an accuracy rate of about 2.8% higher than that of ResNet-50 on the Fer2013 test set. Compared with traditional residual network structure, it verifies that the network proposed in this paper has better effects in the field of facial expression recognition.

针对用于特征提取的ResNet网络结构，还存在过拟合、计算速度慢、准确率提升空间等问题。提出了一种混合深度分离残差网络用于图像特征提取。研究中使用了残差块和逆残差块，使得三种残差单元在训练过程中迭代交互。这一过程促进了深度学习神经网络的特征表示，缓解了传统表情识别研究中由于深度网络层次导致的特征分散问题。实验使用了 CK+ 和 fer2013 数据集。多个实验结果的综合表明，优化后的深度学习模型在 Fer2013 测试集上的准确率比 ResNet-50 高约 2.8%。