统一实验范式下GSR与ECG的短时情感识别 性能差异分析

Performance Difference Analysis of GSR and ECG in Short-Term Emotion Recognition under a Unified Experimental Paradigm

摘要: 当前生理信号情感识别领域已经证实单信号、短时程在技术上是可行的，有力推动了情感实时识别的发展，但是不同信号适用性差异尚不明确。本文在已证实有效的心电(ECG)短时情感识别框架下，系统引入皮肤电(GSR)信号在统一的实验范式下进行对照研究。通过电影片段诱发被试喜、怒、哀、惧和平静5种基本情感，同步采集GSR信号，经去噪、归一化与特征提取后，采用BP神经网络建立分类模型，并与前期研究建立的ECG模型进行对比。结果表明，GSR信号的平均情感识别率为82.29%，低于ECG信号的89.14%；两者在各类情感上的识别率分布也存在明显差异。通过对这种差异的生理机制分析与讨论，本文提出单一生理信号情感识别存在“信号–情感”适用性差异，为实时情感识别系统的信号选型提供了实证依据。两类信号的特征提取与分类总耗时均小于0.15 s，具备良好的实时性。

Abstract: The feasibility of using single physiological signals with short time windows has been demonstrated in the field of emotion recognition, significantly advancing real-time emotion detection. However, the applicability differences among various signal types remain unclear. Under a unified experimental paradigm, this study systematically introduces galvanic skin response (GSR) signals into an already validated short-term electrocardiogram (ECG) emotion recognition framework for a controlled comparison. Five basic emotions—joy, anger, sadness, fear, and calmness—were induced in participants through film clips, while GSR signals were synchronously collected. After denoising, normalization, and feature extraction, a BP neural network model was constructed and compared with the previously established ECG model. The results show that the average recognition rate of GSR signals is 82.29%, which is lower than the 89.14% achieved by ECG signals; the two signal types also exhibit distinct recognition rate distributions across individual emotions. Through physiological mechanism analysis and discussion of these differences, this paper proposes the existence of “signal-emotion” applicability differences in single-signal emotion recognition, providing empirical evidence for signal selection in real-time emotion recognition systems. The total time required for feature extraction and classification of both signal types is less than 0.15 seconds, demonstrating good real-time performance.