现实场景下基于数据增强的人脸表情识别

Facial expression recognition based on data augmentation in the real scenes

现实场景下获取的人脸表情图像受姿态、背景、性别、种族等的影响,使得在此类不可控环境下的人脸表情识别率低。针对上述问题,现在出现了各种数据增强的方式,用于实现现实场景下的人脸表情图像数据增强,目的是通过增加数据集的多样性来提升分类精确度。实验结果证明,文中所提方法相较于传统的将所有表情进行同步数据增强的方法,在FER2013数据集上实现了50%识别精确度的提升,且将损失维持在了1.5左右。

Because the facial expression images acquired in the real scene have the influence of posture,background,gender,race,etc.,the recognition rate of facial expressions in such an uncontrollable environment is low.In response to the above problems,various data augmentation methods have emerged to achieve facial expression image data augmentation in real scenes.The aiming is to improve the facial expression classification accuracy by increasing the diversity of the dataset.Experimental results show that compared with the traditional method of synchronous data enhancement of all expressions,the proposed method achieves a 50%improvement in recognition accuracy on the FER2013 dataset,and the loss is maintained at about 1.5.