A Hybrid DenseNet-SVM Approach for EEG-Based Emotion Recognition
Abstract
Emotion recognition based on electroencephalogram (EEG) signals is a challenging task due to the nonlinear nature and low spatial resolution of brain activity. In the context of deep learning, highly complex architectures often face the risk of hidden overfitting when dealing with noisy signals. In this work, we propose a hybrid architecture that combines the statistical robustness of Differential Entropy (DE) with a customized DenseNet-121 network for spatial feature extraction. The proposed methodology involves mapping DE features onto the 9 × 9 grid to preserve cortical topology, followed by a fine-tuning process limited to 50 epochs as a regularization strategy. Instead of a standard softmax layer, a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel is used to maximize the classification margin. This model was evaluated on the SEED dataset, which consists of EEG signals collected while participants viewed emotional film clips as visual stimuli designed to elicit positive, neutral, and negative emotions. Experimental results using the SEED dataset, under a trial-based cross-validation scheme, achieved an average accuracy of 95.22% with a low standard deviation of 0.22%. These results demonstrate that the proposed hybrid approach provides stable and reliable performance.