TSCL-LwF: A Cross-Subject Emotion Recognition Model via Multi-Scale CNN and Incremental Learning Strategy

Area of Science:

• Affective Computing
• Human-Computer Interaction
• Biomedical Signal Processing

Background:

• Wearable affective human-computer interaction requires comfortable, sparse-channel electroencephalography (EEG) signals.
• Limited EEG information and inter-subject variability pose challenges for cross-subject emotion recognition.

Purpose of the Study:

• To develop a robust cross-subject emotion recognition model using sparse-channel EEG.
• To address the limitations of sparse data and individual differences in emotion recognition.

Main Methods:

• Proposed TSCL-LwF model combining a multi-scale convolutional network (TSCL) and Learning without Forgetting (LwF).
• TSCL extracts spatio-temporal characteristics from sparse-channel EEG.
• LwF strategy adapts to target domain data while retaining source domain knowledge.

Main Results:

• Achieved 77.26% accuracy for valence and 80.12% for arousal classification on the DEAP dataset.
• Demonstrated superior performance on a self-collected EPPVR dataset.

Conclusions:

• Successful cross-subject emotion recognition using sparse-channel EEG is feasible.
• This facilitates the development of practical wearable EEG technologies for emotion recognition.