For emotion classification in facial expression recognition (FER), the performance of both traditional statistical methods and state-of-the-art deep learning methods are highly dependent on the quality of data. Traditional methods use image preprocessing (such as smoothing and segmentation) to improve image quality. However, the results still fail to meet the quality requirements of the emotion classifiers in FER. To address the above issues, this paper proposed a novel framework based on reinforcement learning for pre-selecting useful images(RLPS) for emotion classification in FER, which is made up of two modules: image selector and rough emotion classifier. Image selector is used to select useful images for emotion classification through reinforcement strategy and rough emotion classifier acts as a teacher to train image selector. Our framework improves classification performance by improving the quality of the dataset and can be applied to any classifier. Experiment results on RAF-DB, ExpW, and FER2013 datasets show that the proposed strategy achieves consistent improvements compared with the state-of-the-art emotion classification methods in FER.¹

对于面部表情识别（FER）中的情感分类，传统的统计方法和最新的深度学习方法的性能都高度依赖于数据质量。传统方法使用图像预处理（例如平滑和分割）来提高图像质量。但是，结果仍然不能满足FER中情感分类器的质量要求。针对上述问题，本文提出了一种基于强化学习的框架，用于在FER中预先选择有用的图像进行情感分类，该框架由图像选择器和粗糙情感分类器两个模块组成。图像选择器用于通过强化策略来选择用于情感分类的有用图像，粗略的情感分类器充当训练图像选择器的老师。我们的框架通过提高数据集的质量来提高分类性能，并且可以应用于任何分类器。在RAF-DB，ExpW和FER2013数据集上的实验结果表明，与FER中最新的情感分类方法相比，该策略取得了一致的改进。^1个