Abstract:

Local occlusion often occurs in face images in practical applications, which will result in lower recognition rate and robustness. Therefore, in view of the current situation, an expression recognition model based on improved Generative Adversarial Network（GAN） is proposed. Firstly, the occluded face image is filled and repaired by the antagonistic learning of the generator composed of automatic encoder and two discriminators（local discriminator and global discriminator）. Then, after the global discriminator, multi-classification layers are added to make use of the part of the global discriminator. The expression classifier is composed of convolution layer and multi-classification layer. Finally, through doing experiments, comparing the recognition rates of the expression and different algorithms between facial images with different occlusion areas before and after the occlusion area is filled. The experimental results show that the recognition rate will be higher, especially the recognition rate of large area occlusion.

Key words: Generative Adversarial Network（GAN）, occluding face, face filling, convolutional neural network

摘要：

针对实际应用中人脸图像存在局部遮挡的情况经常发生，会造成识别率下降和鲁棒性降低。因此针对目前存在的这种情况，提出一种基于改进生成式对抗网络（Generative Adversarial Network，GAN）的表情识别模型，先利用由自动编码器构成的生成器和两个鉴别器（局部鉴别器和全局鉴别器）的对抗学习对遮挡人脸图像填补修复，再在全局鉴别器后面添加多分类层，利用全局鉴别器的部分卷积层并在后面添加多分类层构成表情分类器进行表情识别。最后通过实验进行了不同遮挡面积的人脸图像在填补前后表情识别率的对比和不同算法的识别率对比，实验结果证明识别率会更高，尤其提高了人脸大面积遮挡的识别率。

关键词: 生成对抗网络, 遮挡人脸, 人脸填补, 卷积神经网络