MFENet: A Multi-Feature Extraction Network for Enhanced Emotion Detection Using EEG and STFT

Developing computationally efficient models for EEG-based emotion recognition is essential for enabling scalable and responsive brain-computer interface (BCI) systems. However, the inherent nonstationarity of EEG signals and individual variability across subjects complicate the development of models that generalize reliably across diverse affective states. To address these challenges, this paper introduces MFENet, a lightweight convolutional neural network tailored for emotion recognition from electroencephalogram (EEG) data. The proposed approach leverages short-time Fourier transform (STFT) to convert one-dimensional EEG signals into two-dimensional spectrograms, thereby capturing rich time-frequency information while preserving signal structure. MFENet employs multi-feature extraction blocks and grouped convolutions to enhance representational diversity without incurring high computational costs. Comparative evaluation on the SEED and SEED-IV datasets demonstrates that MFENet achieves consistently superior performance over GoogLeNet, ResNet-101, and Inception-v3, with improvements of 11.7%, 16.3%, and 8.6% on the SEED dataset, and 4.9%, 2.5%, and 1.4% on the SEED-IV dataset, respectively. These results demonstrate that MFENet achieves an effective balance between classification performance and architectural simplicity, making it well-suited for deployment in real-time, emotion-aware BCI environments.