Analysis of Facial Expressions in Class Based on Lightweight Convolutional Neural Network

Researches show that analyzing students facial expressions in class is contribute to improving the accuracy of learning behavior diagnosis which can effectively improve the quality of teaching. In order to improve the accuracy and real-time of facial expression recognition in class, we design a lightweight convolutional neural network(GEGNet) based on VGG16 framework with novel feature fusion and attention mechanism. At first, channel pruning and layer pruning are used to reduce redundant features, and Ghost convolution is adopted to replace the traditional convolution which can reduce the number of parameters. Secondly, efficient channel attention(ECA) is integrated into the feature extraction process which enhance feature extraction capability. Furthermore, a fusion of deep and shallow features is performed before regression module to improve the recognition accuracy of the model. The experiment accuracy of GEGNet on the homemade dataset reached 94.25% which is better than VGG16 model. In addition, the calculation volume of GEGnet is 8.5% as much as VGG16, and the parameters is 9.9% of VGG16. Our model also achieves the same performance in the comparative experiments on CK+ and FER2013, which furtherly proves the effectiveness of our model.