关键词: 数据增强, 生成对抗网络, 人脸表情识别, 深度学习

Abstract:

Deep learning methods have significantly advanced in facial expression recognition. But, facial expression databases usually do not have enough data. To solve this problem, this paper proposes a static image data augmentation method. A multi-domain image-to-image translation model based on StarGAN is implemented by modifying the reconstruction loss, which can generate multi-expression facial images from the one of a certain expression. Experiments on CK+ expression database show that this method can improve the accuracy and generalization capacity of recognition models, and can be used for reference to solve the problem of data imbalance.

Key words: data augmentation, generative adversarial networks, facial expression recognition, deep learning