To address the limited capability of traditional facial expression recognition methods that rely on spatial-domain features in modeling fine-grained high-frequency texture representations of expressions, this study proposes a spatial-frequency domain fusion and gradient-enhanced facial expression recognition model, SFA-GCE Net. The model uses ResNet18 as the backbone network and introduces a learnable frequency decomposition (LFD) mechanism to adaptively model high-frequency texture information while maintaining efficient spatial-feature extraction. Meanwhile, a dual-domain adaptive fusion module (DAFM) is designed to effectively integrate complementary spatial-domain and frequency-domain features. In addition, a gradient-enhanced cross-entropy loss (GCE Loss) function is introduced during training to improve the model’s discriminative ability on hard samples by selectively enhancing the gradient signals of confusing negative classes. Experimental results show that SFA-GCE Net achieves accuracy improvements of 2.48% and 2.94% over the ResNet18 baseline on the public RAF-DB and CK+ datasets, respectively, while introducing only a limited number of additional parameters and a limited amount of computational overhead. This shows that the proposed method achieves a good balance between performance improvement and model efficiency, demonstrating its application potential in resource-constrained scenarios.