Expression Recognition and Classification Based on Jump-Layer Optimization of Convolutional Neural Network

In multi-layer neural networks, high-level abstract features are extracted by convolutional layers at the end, which lack low-level detailed features. Meanwhile, the requirements of training parameters are also higher because of the deepening of network layers, which increases the difficulty of training and is more prone to the problems of gradient disappearance or explosion. In this paper, an optimized jump-layer convolutional neural network (JCCN) structure is proposed to modify facial expression recognition classification network model. This method effectively combines low-level detailed features and high-level abstract features through jump-layer connections. Gradient disappearance caused by too much network layers and back propagation parameter transfer are modify via the approach. The proposed method can reduce the risk of network training overfitting and enhance the nonlinearity of the data. At the same time, the 1*1 convolution kernel introduced reduces the training cost and the number of parameters effectively. The experimental results show that the network has a good performance on the data sets FER2013 and CK+. It is anticipated that facial expression recognition and classification methods based on convolutional networks would benefit from this paper.