Application of a Hybrid CNN-SE-GCN Model with Multi-Loss Optimization and Adversarial Samples in Real-Time Emotion Recognition

In recent years, with the rapid development of artificial intelligence and internet technologies, facial expression recognition has gained widespread application and attention in fields such as human-computer interaction and intelligent healthcare. However, existing methods still face significant limitations when dealing with small samples, imbalanced datasets, and adversarial attacks. Therefore, we propose a hybrid Convolutional Neural Network (CNN)- Squeeze-and-Excitation (SE) Graph Convolutional Network (GCN) model that integrates multi-loss optimization and adversarial samples for real-time facial expression recognition. The model draws inspiration from the generator-discriminator architecture of Generative Adversarial Networks (GANs). The generator is based on an improved ResNet18, incorporating an SE module to extract local features from images, while leveraging GCNs to process high-dimensional geometric information, thereby effectively fusing global and local features. To enhance the generator's adaptability to adversarial samples, we employ the Fast Gradient Sign Method (FGSM) to generate adversarial samples and combine them with the original data for joint training, further optimizing the generator's feature representation and robustness. Additionally, we designed a geometric information discriminator to distinguish generated features from real geometric information, strengthening the discriminator's capability through adversarial training. Furthermore, we propose a multi-loss optimization strategy that integrates classification loss, geometric constraint loss, and adversarial training loss to optimize the model from multiple dimensions. Experimental results on the CK+ dataset demonstrate that the proposed method achieves high accuracy and robustness in small-sample facial expression recognition tasks, effectively addressing data imbalance and adversarial sample interference. Finally, we developed a web system that applies the proposed method to practical scenarios, providing an efficient solution for real-time facial expression recognition tasks.