Facial Expression Recognition (FER) has garnered significant attention in recent years due to its wide applications in human-computer interaction, psychological analysis, and affective computing. With the advancements in deep learning, FER models have achieved remarkable accuracy on benchmark datasets. However, most existing methods perform well on in-distribution datasets but struggle with out-of-distribution data, revealing weak generalization ability. This limitation hampers their effectiveness in real-world scenarios, where the distribution of test data often differs from that of the training data. To address this challenge, we propose a dual-network framework that integrates a global network and a local network to extract complementary features from images. The extracted features are subsequently fused to enhance the model’s generalization capability. Furthermore, we propose a novel loss function that integrates cross-entropy loss with contrastive loss to optimize model training. The contrastive loss is designed to measure the discrepancy between global and local features, encouraging the model to learn more complementary representations. Ablation experiments validate the effectiveness of the proposed loss function. Experimental results demonstrate that our method outperforms existing approaches on multiple FER datasets, achieving superior generalization to out-of-distribution datasets.