Because of the different features of the same facial expressions from different angles, most of the methods are only suitable for face images, and the accuracy of facial expression recognition is low. Therefore, a multi-angle facial expression recognition method based on generative adversarial networks (GAN) is proposed. Firstly, the depth regression network is used to detect the key points of the face image to achieve face alignment, so as to reduce the difficulty of feature extraction. Then, the image is input to GAN. The generator is composed of encoder and decoder, and a skip connection is designed. In the encoding phase, the generator can unlock the correlation between the facial expression image and the angle, and in the decoding stage, it can generate different angle facial expression images by adding other angle information. Finally, the multi-angle facial expression images are sent to the convolution neural network for classification and learning, in which the loss weight is adjusted dynamically to improve the recognition accuracy by introducing resistance loss, recognition loss, content loss, and center loss. The experimental results on Multi-pose illumination expression (PIE) and celebrities in frontal profile (CFP) datasets show that the performance of the proposed method is the best when the learning rate is 0.0002, and the recognition accuracy under different angles is higher than other comparison methods, so it has practical application significance.

中文翻译：

---

基于生成对抗网络的多角度人脸表情识别

由于相同的面部表情从不同的角度具有不同的特征，大多数方法只适用于人脸图像，面部表情识别的准确率较低。因此，提出了一种基于生成对抗网络（GAN）的多角度面部表情识别方法。首先，利用深度回归网络检测人脸图像的关键点，实现人脸对齐，从而降低特征提取的难度。然后，图像被输入到 GAN。生成器由编码器和解码器组成，并设计了跳跃连接。在编码阶段，生成器可以解锁人脸表情图像与角度的相关性，在解码阶段，可以通过添加其他角度信息来生成不同角度的人脸表情图像。最后，将多角度人脸表情图像送入卷积神经网络进行分类学习，其中通过引入阻力损失、识别损失、内容损失和中心损失，动态调整损失权重以提高识别准确率。在多姿态光照表情（PIE）和正面轮廓（CFP）数据集上的实验结果表明，当学习率为0.0002时，所提方法的性能最好，并且不同角度下的识别准确率高于其他方法。比较方法，因此具有实际应用意义。其中通过引入阻力损失、识别损失、内容损失和中心损失来动态调整损失权重以提高识别准确率。在多姿态光照表情（PIE）和正面轮廓（CFP）数据集上的实验结果表明，当学习率为0.0002时，所提方法的性能最好，并且不同角度下的识别准确率高于其他方法。比较方法，因此具有实际应用意义。其中通过引入阻力损失、识别损失、内容损失和中心损失来动态调整损失权重以提高识别准确率。在多姿态光照表情（PIE）和正面轮廓（CFP）数据集上的实验结果表明，当学习率为0.0002时，所提方法的性能最好，并且不同角度下的识别准确率高于其他方法。比较方法，因此具有实际应用意义。