Facial Expression Recognition Based on Improved Densenet Network

Abstract

As a branch of image recognition, facial expression recognition helps to carry out medical, educational, security and other work more efficiently. This article combines deep learning knowledge and conducts research on expression recognition based on DenseNet121, a dense convolutional neural network that integrates attention mechanisms for multi-scale feature extraction. Firstly, in response to the insufficient ability of DenseNet121 to extract complex facial expression features, multi-scale feature extraction dense blocks were introduced to replace DenseBlocks used to extract features of different sizes; Secondly, using multi-scale feature extraction convolutional blocks to replace the large convolutional kernel at the head of DenseNet121 further enriches feature extraction; Finally, in order to extract more important features from the channel dimension, we consider combining ECA channel attention mechanism to help improve model performance. The experiment proves that the model proposed in this chapter has improved recognition accuracy by 2.034% and 3.031% compared to DenseNet121 on the FER2013 and CK+datasets, respectively. It also has certain advantages compared to other commonly used classification models.