Abstract: To further improve the recognition rate of facial expressions, a face recognition method based on deep learning and Stacked Hybrid Auto-Encoder（SHAE） is adopted. The structure of the method is a 5-layer network structure composed of a Denoising Auto-Encoder（DAE）, a Sparse Auto-Encoder（SAE）, and an Auto-Encoder（AE）. In order to increase the robustness and generalization ability of the network, a DAE is used to extract features from the samples. In order to reduce the dimensions of the extracted features and to extract further abstract sparse features, a SAE is used for cascading, and further processing of features. The training process begins with pre-training and overall fine-tuning of the unlabeled data, initializing and updating the weight of the whole structure, and then testing and training with labeled data. Experiments on two datasets, JAFFE and CK+, show that this method has a better recognition effect than a purely stacked DAE or apurely stacked SAE.

Key words: face expression recognition, Stacked Hybrid Auto-Encoder, Sparse Auto-Encoder, Denoising Auto-Encoder

摘要： 针对进一步提高人脸表情识别率的问题，采用了一种基于深度学习的堆栈式混合自编码器（Stacked Hybrid Auto-Encoder，SHAE）的人脸表情识别方法。该方法的结构是由去噪自编码器（Denoising Auto-Encoder，DAE）、稀疏自编码器（Sparse Auto-Encoder，SAE）以及自编码器（Auto-Encoder，AE）组合而成的5层网络结构。为了增加网络的鲁棒性以及泛化能力，采用去噪自编码器对样本进行提取特征，为了对提取的特征进行降维以及进一步提取更抽象的稀疏特征，采用稀疏自编码器进行级联，来对特征进一步处理。训练过程首先由无标签的数据进行预训练和整体微调，对整个结构的权重进行初始化和更新调整，然后使用有标签的数据进行测试训练。在JAFFE和CK+两个数据集上实验显示，相较于单纯的堆栈式去噪自编码或者单纯的堆栈式稀疏自编码，该方法具有更好的识别效果。

关键词: 人脸表情识别, 堆栈式混合自编码器（SHAE）, 稀疏自编码器（SAE）, 去噪自编码器（DAE）