Abstract
As human-computer interaction technology evolves, emotional computing (EC) is becoming increasingly significant within the Internet of Things (IoT). This paper presents a high-accuracy emotion recognition system utilizing Double Tree Complex Wavelet Transform (DTCWT) and Support Vector Machines (SVM). By analyzing EEG signals from 16 subjects induced by video stimuli, the proposed method achieves a 90.61% classification accuracy across three emotional states: calm, happy, and sad. The use of DTCWT effectively addresses common signal processing issues like frequency aliasing and shift sensitivity, providing a robust framework for EC in IoT applications.
Introduction
Emotional Computing (EC) aims to enable machines to recognize and respond to human emotions, a critical component for seamless human-computer interaction in IoT environments. While electroencephalogram (EEG) signals provide direct insight into emotional states, they are complex and prone to noise. Traditional methods like the Discrete Wavelet Transform (DWT) often suffer from translation sensitivity and frequency aliasing, which degrade feature extraction quality. This study introduces DTCWT to overcome these limitations and improve recognition reliability.
Methodology
The core of the proposed system is the Double Tree Complex Wavelet Transform (DTCWT). Unlike the standard DWT, DTCWT employs two parallel filter banks for the real and imaginary parts of the transformation. This architecture ensures shift invariance and significantly reduces frequency aliasing, leading to cleaner feature vectors.
Experimental Protocol
Data was collected from 16 healthy undergraduate subjects using a 64-lead Neuroscan device. The experimental procedure involved specific intervals of stimulation and rest to capture distinct emotional responses.
Signal Processing and Feature Extraction
Raw EEG data was band-pass filtered and downsampled to 125 Hz. The signals were then decomposed into various frequency bands ($\theta, \delta, \alpha, \beta$) using DTCWT. The reconstruction of these bands allows for precise feature extraction.
Key features extracted included the $\beta$-band phase, $\alpha$-band power, and the $\beta/\theta$ power ratio. These features were fed into an SVM classifier structured as a "two-fork tree" to handle multi-class emotion sorting.
Results and Discussion
The system demonstrated exceptional performance, achieving a 90.61% accuracy in distinguishing between calm, happy, and sad states. A significant finding was the sensitivity of the instantaneous phase change in the $\beta$ band, which serves as a reliable marker for emotional shifts.
The use of DTCWT proved superior to DWT by maintaining frequency components strictly within their theoretical bands, preventing the "leakage" or aliasing that often confuses machine learning models. Furthermore, normalizing data by subtracting individual "calm" baselines proved essential for cross-subject reliability.
Conclusion
This research highlights the effectiveness of DTCWT combined with SVM for robust emotion recognition in the IoT domain. By solving fundamental signal processing challenges, the proposed method offers a high-accuracy solution for real-time emotional monitoring. Future work will explore deep learning architectures to further automate and refine feature discovery from complex wavelet coefficients.
Summary updated