An Expression Recognition Method based on Improved Convolutional Network

In facial expression recognition tasks, traditional machine learning methods are based on manual feature extraction, which has high time complexity and poor stability. However, the existing convolutional network model has some problems such as insufficient feature extraction and large number of parameters. To solve these problems, this paper adds deep detachable convolutional network into the ordinary convolutional layer to reduce the number of parameters and operation cost. We add Squeeze - and - Excitation block to the outputs of the deeply separable convolution layer to reallocate the weights of the output channels according to the importance. What’s more, we introduce the residual mechanism in Resnet network to reduce the disappearance of gradient. The designed model was trained in Fer2013 data set. The nerual network proposed in this paper achieves 70% accuracy on the test set after training when the number of parameters is much smaller than many traditional convolutional networks, and achieves more accurate and lightweight facial expression recognition.