ICTD: Combination of Improved CNN-Transformer and Enhanced Deep Canonical Correlation Analysis for Eye-Movement...

Area of Science:

• Affective computing
• Human-computer interaction
• Biomedical signal processing

Background:

• Emotion classification using eye-movement features is common due to ease of data acquisition and ocular response correlation.
• Existing methods struggle with weak feature-emotion correlations and lack of prominent key feature presentation.

Purpose of the Study:

• To address limitations in current emotion recognition methods by proposing an improved CNN-Transformer with enhanced deep canonical correlation analysis (ICTD).
• To enhance feature representation and classification performance for more accurate emotion recognition.

Main Methods:

• Preprocessing and reconstruction of raw eye-movement signals for informative feature extraction.
• Utilizing Convolutional Neural Networks (CNNs) for local features and Transformer architectures for global features.
• Incorporating an incremental feature feedforward network to enhance the Transformer's focus on salient information.
• Employing deep canonical correlation analysis based on cosine similarity for classification.

Main Results:

• The ICTD framework demonstrated superior performance over baseline approaches on SEED-IV, SEED-V, and eSEE-d datasets.
• Achieved 81.8% arousal and 85.2% valence classification on the eSEE-d dataset.
• Reached 91.2% for four-category emotion classification on SEED-IV and 85.1% for five-category classification on SEED-V.

Conclusions:

• The proposed ICTD framework effectively enhances feature representation and classification accuracy in emotion recognition.
• Demonstrates significant potential for practical applications in emotion recognition and physiological signal analysis.