A Case Study of EmoNeXt-Tiny Without Self-Attention for Facial Emotion Recognition

Facial emotion recognition (FER) on FER2013 remains challenging due to low-resolution grayscale images, class imbalance, and limited data diversity. This study presents a controlled Tiny-scale ablation of EmoNeXt-Tiny to isolate the contribution of the self-attention (SA) regularization term. Specifically, SA regularization is removed from the objective function, while all other components and conditions are preserved, including the ConvNeXt-Tiny backbone, the Spatial Transformer Network (STN), Squeeze-and-Excitation (SE) modules, the preprocessing pipeline, and the training/evaluation protocol. Experiments are conducted on FER2013 using the official train/validation/test split. Under an identical setup, the SA-free variant achieves 72.95% top-1 test accuracy, compared with 73.34% for the full EmoNeXt-Tiny and 71.99% for the ConvNeXt-Tiny baseline. Because the present study reports a single controlled run per configuration (i.e., without multi-seed repetition), the 0.39 percentage-point gap should be interpreted as a preliminary observation that may fall within random variance. Within this constraint, the findings indicate diminishing returns from SA regularization in this Tiny regime once geometric alignment and channel re-weighting are already incorporated. In addition, the SA-free model simplifies optimization by removing an auxiliary loss component and its associated tuning burden. Overall, an STN+SE-enhanced ConvNeXt-Tiny without SA offers a practical accuracy–complexity trade-off for resource-constrained FER applications.