Emotion recognition plays a critical role in mental health monitoring, decision-making, and enhancing human–computer interaction. The complex, non-stationary, and multi-component nature of EEG signals presents significant challenges for accurately detecting emotional states. Traditional time–frequency (TF) methods often fail to capture rapid and transient oscillatory events in EEG signals due to limited resolution. Moreover, the lack of standardized evaluation protocols in the field has led to inflated accuracy estimates, particularly when spectrogram-level data splitting allows correlated electrode samples from the same trial to appear in both training and test sets. To address these limitations, this study presents a rigorous evaluation of the Super-Resolution Superlet Transform (SLT) against both conventional TF methods (STFT, CWT, SPWVD) and adaptive signal decomposition methods (EMD-HHT, VMD-HHT) for EEG-based emotion recognition, using a self-attention convolutional neural network (SA-CNN). All experiments employ strict trial-level cross-validation to eliminate data leakage. The framework is evaluated on the DEAP and DREAMER datasets for subject-dependent (SD) and subject-independent (SI) classifications of arousal and valence states. Under the corrected evaluation protocol, SD classification achieves accuracies of 74.41% (arousal) and 72.04% (valence) on DEAP, and 78.63% (arousal) and 76.82% (valence) on DREAMER. For SI classification, it attains 70.86% (arousal) and 71.36% (valence) on DEAP, and 73.51% (arousal) and 75.46% (valence) on DREAMER. SLT consistently outperforms all five baseline TF methods across both evaluation protocols. Ablation studies confirm the contribution of the self-attention layer, and Grad-CAM analysis demonstrates that self-attention selectively focuses on neuroscience-consistent frequency bands. • First comparison of SLT vs. adaptive decomposition for EEG emotion recognition. • Electrode-level data leakage corrected; trial-level CV reduces SD accuracy 15–23pp. • SLT outperforms all five baseline TFRs under leak-free evaluation on DEAP/DREAMER. • Grad-CAM shows self-attention targets β / γ for arousal and α bands for valence. • Comprehensive SD and SI evaluation with Friedman/Nemenyi statistical testing.
Kumar et al. (Wed,) studied this question.