Abstract: At present, most existing facial expression recognition methods assume that each facial expression in the training set corresponds to a single emotion, which is typically cast as a classification problem. However, in actual situations, a face expression rarely expresses pure emotion, but often a mixture of different emotions. Therefore, samples of similar expression have a correlation at the emotional level, and this correlation also often leads to ambiguity in the sample??s expression labels. Namely, each emotion sample is associated with a latent label distribution. For this reason, this paper proposes a totally data-driven label distribution learning approach to adaptively learn the latent label distributions. Without any pre-hypothesized label distribution, the relationship between each expression and its multiple emotions can be obtained. This method can get the specific description of each emotion contained in the expression and the mapping of the expression image to the emotion distribution. Experimental results show that the proposed method has high accuracy in facial expression recognition and can effectively deal with the facial emotion analysis problem.

Key words: facial expression recognition, label distribution, subspace learning, description degree

摘要： 现有的大多数面部表情识别方法都是假定样本中每个人脸表情对应单一的情绪，而后作为分类问题进行解决。但是在实际情况中，一个人脸表情往往是多种不同基础情绪的混合体。因此，具有相似表情的样本在基础情绪层面存在一定的相关性，这种相关性也通常会导致样本的表情标签呈现多样性。也就是说，每个样本的表情状况与潜在的情绪标签分布相关联。为此，提出了一种通过数据进行自适应学习潜在标签分布的方法，不需要任何预先假设的标签分布形式，便可得到每个表情与其相应的多个情绪的关联情况。该方法可得到每个表情所包含情绪的特定描述度以及表情图像到情绪分布的映射。实验结果表明，该方法在表情识别上具有较高准确率，且能够有效地解决人脸表情的分析问题。

关键词: 面部识别, 标签分布, 子空间学习, 描述度