基于局部空间特征引导的表情识别算法

Expression recognition algorithm guided by local spatial features

摘要:

面部表情识别在计算机视觉任务中受到越来越多的关注，由于真实场景中的表情往往包含着大量由姿态、年龄、图像质量、标注等因素带来的噪声，大大增加了类内变化，给表情的分类任务带来了很大的困难。现有的基于此类问题的研究往往聚焦于数据本身，通过对数据进行筛选或者扩大模型接受的数据类型的形式提高识别能力，没有考虑到卷积网络本身对图像特征关注的局限性。针对该问题，提出了一种基于局部空间特征引导的卷积神经网络，对于特征图的某部分像素点进行强调，引导卷积网络的深层特征图能够关注到多个对分类有效的局部面部区域，同时使用对数据重标记的形式抑制由标签错误导致的噪声问题。经过在多个公开的表情识别数据集中测试，并与多个同类方法对比，所提方法具有较好的识别效果。

Abstract:

Facial expression recognition has received increasing attention in computer vision tasks. In real-world scenarios, facial expressions often contain a significant amount of noise introduced by factors such as pose, age, image quality, and annotation, which have greatly increased intra-class variation and have posed significant challenges for facial expression classification tasks. The existing researches addressing this problem often focus on the data itself, improving recognition capabilities by filtering or expanding the types of data accepted by the models, without considering the limitations of the convolutional networks in attending to image features. To address this issue, this paper proposed a convolutional neural network（CNN） based on local spatial feature guidance. It emphasizes certain pixels in the feature maps, enabling deep layers of the convolutional network to attend to multiple local facial regions that are effective for classification. Additionally, a re-labeling approach was employed to suppress noise caused by label errors. The proposed method was tested on multiple publicly available facial expression recognition datasets and has achieved better recognition performance compared to several existing methods.