微表情是一种持续时间短、强度低的自发性面部表情，因为其能够客观性揭示真情实感的特点而在很多领域都有着重要作用，引起广泛研究。微表情识别早期是基于传统机器学习方法进行研究，后来随着深度学习广泛运用于图像识别，研究人员将其运用于微表情识别中。但由于公开的自发微表情数据集少原因，传统神经网络并没有取得较好的提升效果。近年来，胶囊网络在处理微型数据集上取得成功，为提高深度学习在微表情识别中的表示能力，本文采用了一种基于胶囊网络的微表情识别方法。
针对目前比较流行的公开微表情数据集SMIC、CASME II、SAMM，我们首先引入了一种寻找微表情序列顶点帧的方法，将三个数据集整合成。接着构建由输入层、卷积层、主要胶囊层、输出胶囊层组成的用于微表情识别的胶囊网络，并介绍了胶囊网络中连接各胶囊层的动态路由机制。最终实验方法的有效性是通过使用LOSO交叉验证方法在跨数据库微表情基准上进行评估的。实验结果显示，我们的方法在性能上优于基准方法（LBP-TOP）和其他最先进的卷积神经网络模型。

Micro-expressions, which are brief and low-intensity spontaneous facial expressions, have gained significant attention in various domains due to their ability to objectively reveal genuine emotions. Early research on micro-expression recognition primarily employed traditional machine learning methods. However, with the widespread adoption of deep learning in image recognition, researchers started exploring the application of deep learning techniques in micro-expression recognition. Nonetheless, the limited availability of publicly accessible datasets containing spontaneous micro-expressions posed challenges to achieving substantial improvements using traditional neural networks. In recent years, capsule networks have achieved success in handling small-scale datasets. To enhance the representation capabilities of deep learning in micro-expression recognition, this paper adopts a micro-expression recognition method based on capsule networks.

To address the scarcity of publicly available micro-expression datasets, particularly SMIC, CASME II, and SAMM, we introduce a method for identifying key frames in micro-expression sequences, enabling the integration of these three datasets. Furthermore, we construct a Capsule Network for micro-expression recognition, comprising an input layer, convolutional layers, primary capsule layer, and output capsule layer. Additionally, we introduce a dynamic routing mechanism for connecting the capsule layers within the network. The effectiveness of our proposed method is evaluated using the Leave-One-Subject-Out (LOSO) cross-validation technique on a cross-database micro-expression benchmark. Experimental results demonstrate that our method surpasses baseline approaches (such as LBP-TOP) and other state-of-the-art convolutional neural network models in terms of performance.Micro-expressions, which are brief and low-intensity spontaneous facial expressions, have gained significant attention in various domains due to their ability to objectively reveal genuine emotions. Early research on micro-expression recognition primarily employed traditional machine learning methods. However, with the widespread adoption of deep learning in image recognition, researchers started exploring the application of deep learning techniques in micro-expression recognition. Nonetheless, the limited availability of publicly accessible datasets containing spontaneous micro-expressions posed challenges to achieving substantial improvements using traditional neural networks. In recent years, capsule networks have achieved success in handling small-scale datasets. To enhance the representation capabilities of deep learning in micro-expression recognition, this paper adopts a micro-expression recognition method based on capsule networks.
To address the scarcity of publicly available micro-expression datasets, particularly SMIC, CASME II, and SAMM, we introduce a method for identifying key frames in micro-expression sequences, enabling the integration of these three datasets. Furthermore, we construct a Capsule Network for micro-expression recognition, comprising an input layer, convolutional layers, primary capsule layer, and output capsule layer. Additionally, we introduce a dynamic routing mechanism for connecting the capsule layers within the network. The effectiveness of our proposed method is evaluated using the Leave-One-Subject-Out (LOSO) cross-validation technique on a cross-database micro-expression benchmark. Experimental results demonstrate that our method surpasses baseline approaches (such as LBP-TOP) and other state-of-the-art convolutional neural network models in terms of performance.