关键词: 表情识别, 卷积神经网络, 随机权重平均, 随机梯度下降法, Fer2013数据集, 网络级联

Abstract:

In order to improve the detection accuracy of expression recognition technology, a face expression recognition method based on cascade network optimized by SWA is proposed. Compared with a single convolutional network, multi-network cascading reachs higher detection accuracy. With respect to the traditional training method SGD, SWA training method enhances the generalization ability of the sub-network in the cascade network, which further improves the overall performance of the model. By testing on the Fer2013 dataset, the experimental results shows that the detection accuracy of the network cascaded by the way of weighted summation based on SWA training method reachs 74.478%, which is 1.4% higher than the single network model trained by the traditional SGD method. In addition, the improved model proposed in the present study reachs a higher recognition accuracy than other typical methods.

Key words: expression recognition, convolutional neural network, stochastic weight averaging, stochastic gradient descent, Fer2013 dataset, network cascade