Facial expression recognition has become an increasingly important research topic in pattern recognition and affective computing. Most of facial expression recognition methods assume that the training and testing data come from the same corpus. However, in practical situations, this assumption does not hold, in which the data are often collected from different scenarios, e.g., different races, environments, or devices, and the recognition performance will drop significantly. To tackle this problem, in this paper, we propose a novel transfer learning method, called latent sparse transfer subspace learning (LSTSL), for cross-corpus facial expression recognition. Specifically, we aim to learn a common subspace in which the target samples can be linearly represented by a few source samples. By introducing an ${\ell }_{2,1}$-norm on the reconstructive transformation matrix, the most discriminative features can be well selected. To guide the new representation learning, we design a novel graph in which the local structure information can be well preserved. Furthermore, we introduce the popular distance metric, i.e., maximum mean discrepancy (MMD), to boost the transfer ability. We conduct extensive cross-corpus experiments on four popular facial expression datasets. The results show that our proposed method can outperform several state-of-the-art transfer learning methods for cross-corpus facial expression recognition.

面部表情识别已成为模式识别和情感计算中越来越重要的研究课题。大多数面部表情识别方法都假设训练和测试数据来自同一个语料库。然而，在实际情况下，这种假设并不成立，其中数据往往来自不同的场景，例如不同的种族、环境或设备，识别性能会显着下降。为了解决这个问题，在本文中，我们提出了一种新的转移学习方法，称为潜在稀疏转移子空间学习（LSTSL），用于跨语料库面部表情识别。具体来说，我们的目标是学习一个公共子空间，其中目标样本可以由几个源样本线性表示。通过引入一个${\ell }_{2,1}$-norm 在重构变换矩阵上，可以很好地选择最具辨别力的特征。为了指导新的表示学习，我们设计了一个新的图，其中可以很好地保留局部结构信息。此外，我们引入了流行的距离度量，即最大平均差异 (MMD)，以提高传输能力。我们对四个流行的面部表情数据集进行了广泛的跨语料库实验。结果表明，我们提出的方法可以胜过几种用于跨语料库面部表情识别的最先进的迁移学习方法。