Improved Resnet18 Expression Recognition Algorithm

Expression recognition is a sub-task of computer vision. The accuracy of expression recognition accuracy is an important research content. However, Because facial expressions have different complexity and diversity at different times, expression recognition has always been a research difficulty. Aiming at the problem of feature extraction in expression recognition, to improve the accuracy of expression recognition, this paper proposes an expression recognition algorithm that improves the Resnet18 network model. First, in the Resnet18 network, The main branch of the residual module is embedded with the SEnet attention module. The feature layer is extracted from the channel dimension. Through the adaptive global average pooling method, the feature layer is compressed, and only the channel dimension information is taken. Then two full connection layers are used to pay attention to the channel information. Finally, the original feature layer is weighted. This method improves the accuracy of image classification tasks, Secondly, the Swish activation function is used to prevent saturation when the gradient gradually approaches zero during slow training, improve the performance of the whole neural network, and improve the accuracy of the network in classification. The experimental results show that, after improvement, compared with the original Resnet18 network, the accuracy on the public data set RAF-DB is improved by 0.59%, reaching 83.64%. Compared with the relevant classification methods, the detection accuracy is effectively improved.