Compound Facial Expression Recognition Based on Graph Convolutional Multi-Label Learning

Composite facial expression recognition is an emerging area within facial expression analysis, addressing the limitations of traditional methods that focus only on basic expressions. This paper proposes a multi-label learning model based on Graph Convolutional Networks to enhance the recognition of complex composite expressions by effectively capturing relationships between expressions. The proposed model consists of two branches: an image feature learning branch and a label relation extraction branch. In the image feature learning branch, facial features are extracted using a pre-trained CNN model, and interdependencies are learned by reconstructing the GCN node feature matrix, integrated with a multilayer perceptron (MLP). The label relation extraction branch constructs a graph of expression category relations, utilizing GCN to model dependencies between expression categories. To address data imbalance, weights are normalized. Experiments show that the proposed method achieves a mean Average Precision (mAP) of 82.37%, an OF1 score of 77.34%, and a CF1 score of 75.19% on the RAF-DB dataset, and an mAP of 86.53%, an OF1 score of 87.32%, and a CF1 score of 86.57% on the EmotioNet dataset. This method effectively captures deep relationships among composite expressions, significantly improving classification performance and providing strong technical support for composite facial expression recognition.