基于生成对抗网络的遮挡表情识别

针对实际应用中局部遮挡会影响人脸表情识别，提出一种基于生成对抗网络（GAN）的表情识别算法。先对遮挡人脸图像填补修复，再进行表情识别。其中GAN的生成器由卷积自动编码机构成，与鉴别器的对抗学习使得生成的人脸图像更加逼真；由卷积神经网络构成的鉴别器具有良好的特征提取能力，添加多分类层构成了表情分类器，避免了重新计算图像特征。为了解决训练样本不足的问题，将CelebA人脸数据集用于训练人脸填补修复，同时表情分类器的特征提取部分得到了预训练。在CK+数据集上的实验证明，填补后的人脸图像真实连贯，并取得了较高的表情识别率，尤其提高了人脸大面积遮挡的识别率。

Occluded facial expression recognition based on generative adversarial networks

Aiming at the fact that partial occlusion affected facial expression recognition in practical applications, this paper proposed an expression recognition method based on generative adversarial networks(GAN). Firstly, this method filled and repaired the occlusion face images, and then performed the expression recognition. The generator of GAN was composed of a convolutional auto-encoder, the face images generated by adversarial learning between generator and discriminator were more vivid. The discriminator was composed of the convolutional neural network, which had good feature extraction ability. It added a multi-classification layer to construct the expression classifier, which avoided feature re-calculation. In order to solve the problem of insufficient training samples, this paper used the CelebA face dataset to train face filling and repairing, and pre-trained the feature extraction part of the expression classifier. Experiments on the CK+ dataset show that the face images after filling are real and coherent, and achieves a higher expression recognition rate. Especially it improves the recognition rate of large-area occlusion of the face.