Emotion recognition is a very important technique for effective interaction between human and artificial intelligence (AI) system. For a long time, facial expression-based methods have been actively studied, and they are showing high recognition performance thanks to powerful deep learning recently. On the other hand, the images of the datasets used in the conventional emotion recognition studies are usually short in length and often generated through intentional expression. Also, continuous domain annotation of emotional labels in dataset configuration requires high cost. In order to overcome such problems, this paper proposes an emotion recognition method based on semi-supervised learning that utilizes an appropriate amount of unlabeled dataset in parallel while minimizing the use of labeled dataset requiring high training cost. The proposed emotion recognition method is based on CNN-LSTM-based regressor for regressing arousal and valence in continuous domain. In addition, we present scenarios and design criteria in which semi-supervised learning can be effectively applied to emotion recognition tasks through experiments using well-known MAHNOB-HCI and AFEW-VA datasets.

情感识别是人类与人工智能（AI）系统之间进行有效交互的一项非常重要的技术。长期以来，基于面部表情的方法已经得到了积极的研究，并且由于最近强大的深度学习，它们表现出很高的识别性能。另一方面，常规情绪识别研究中使用的数据集的图像通常长度较短，并且通常是通过有意表达而生成的。而且，在数据集配置中对情感标签进行连续域注释需要很高的成本。为了克服这些问题，本文提出了一种基于半监督学习的情感识别方法，该方法并行利用适当数量的未标记数据集，同时最大程度地减少需要高训练成本的标记数据集的使用。所提出的情绪识别方法基于基于CNN-LSTM的回归器，用于在连续域中回归唤醒和化合价。此外，我们提出了一些场景和设计标准，其中可以通过使用著名的MAHNOB-HCI和AFEW-VA数据集的实验将半监督学习有效地应用于情感识别任务。