Facial Expression Recognition Method Based on Sparse Batch Normalization CNN

In order to overcome the problem that the convolution neural network facial expression recognition model training is prone to gradient disappear or gradient explode and the fully connected layers with highly complex connection leads to the over-fitting. A sparse batch normalization convolution neural network model is proposed. In the early convolution process, the larger convolution kernels are used continuously to integrate the facial expression features. Batch normalization is applied to the middle layer of the model to adjust the data size and distribution so that the model can be trained quickly. At the fully connected layers, a low retention ratio Dropout connection forms reduce model calculations and over-fitting risk. The performance of the model is validated on the JAFFE dataset and the CK+ dataset, respectively, and the recognition accuracy is over 95%. The model has better recognition performance than the common deep learning model and traditional methods.