| 中文题名： | 用于面孔身份识别的深度神经网络对面孔表情的识别与加工 |
| 姓名： | |
| 保密级别： | 公开 |
| 论文语种： | 中文 |
| 学科代码： | 071101 |
| 学科专业： | |
| 学生类型： | 学士 |
| 学位： | 理学学士 |
| 学位年度： | 2020 |
| 学校： | 北京师范大学 |
| 校区： | |
| 学院： | |
| 第一导师姓名： | |
| 第一导师单位： | |
| 提交日期： | 2020-06-06 |
| 答辩日期： | 2020-05-20 |
| 外文题名： | THE FACIAL EXPRESSION RECOGNITION AND PROCESS OF A DEEP NEUAL NETWORK THAT TRAINED TO PROCESS FACIAL IDENTITY |
| 中文关键词： | |
| 外文关键词： | Face perception ; Facial identity recognition ; Facial expression recognition ; Deep neural network ; VGG-Face |
| 中文摘要： |
人类可以对一张面孔同时进行身份识别和表情识别。
﹀
 面孔身份识别与表情识别是否共用相同的表征？最近，深度神经网络已经可以在面孔身份识别任务上达到人类水平，并且能够从训练任务上实现面孔身份识别和表情识别的分离。那么对于一个深度神经网络来说，在面孔身份识别任务训练中学习到的表征，是否能够支持它完成面孔表情的识别任务？我们使用经过面孔身份识别预训练的网络（VGG-Face），对面孔表情图片刺激库（KDEF）进行迁移学习训练。结果表明，VGG-Face对表情识别迁移学习的平均正确率为92.46%，显著高于随机水平。为了进一步探究用于面孔表情识别的关键表征的所在层级，我们对前几层分别进行了迁移学习测试。结果发现，第一层就已经能够达到50%左右的正确率，在第2至7层的正确率达到了85%~95%。这说明，面孔身份识别的深度神经网络能够进行表情识别，且第二层就已经加工了识别表情所需的关键特征。这为深度神经网络中面部表情识别与面孔身份识别共用相同的表征提供了证据。这暗示着，在人类身上，面孔表情识别与身份识别也很可能共用了相同的知觉特征。 |
| 外文摘要： |
Human can easily recognize both the expression and identity from one face. Do facial identity recognition and facial expression recognition share the representation during the facial perception? Lately, the facial identity recognition ability of deep neural networks have already achieve human level. We can separate facial identity recognition from facial expression recognition in training task. So, for a deep neural network, which was trained only to finish human facial identity recognition task, can the representation that it learned during this training support it finish the facial expression recognition task?We use a deep neural network that was pre-trained to recognize facial identity recognition(VGG-Face), and transfer its high level features to learn facial expression based on KDEF database. The results show that, the average accuracy rate is 92.46%, which is significantly above chance level. To further explore the critical level that process representation used in facial expression recognition, we use transfer learning to test the facial expression ability of different neural levels. The result shows that the accuracy of the first layer has already reached 50%, and the accuracies of layer two to layer seven are 85%~95%. This demonstrate that facial identity deep neural network can recognize facial expression, and the second layer process the critical features used for facial expression recognition. The result provide evidence that facial expression recognition and facial identity recognition share the same representation in deep neural network. This might suggest that, in human, facial expression recognition and facial identity recognition share the same representation.
﹀
|
| 参考文献总数： | 49 |
| 馆藏号： | 本071101/20058 |
| 开放日期： | 2021-06-06 |
谷歌
