Emotion Recognition Based on DEAP Database Physiological Signals

Recognizing and accurately classifying human emotion is a complex and challenging task. Recently, great attention is paid to the emotion recognition methods using three different approaches: based on non-physiological signals (like speech and facial expression), based on physiological signals or based on hybrid approaches. Non-physiological signals are easily controlled by the individual, so these approaches have downsides in real world applications. In this paper, an approach based on physiological signals which cannot be willingly influenced (electroencephalogram, heartrate, respiration, galvanic skin response, electromyography, body temperature) is presented. Publicly available DEAP database was used for the binary classification (high vs. low) considering four frequently used emotional parameters (arousal, valence, liking and dominance). We have extracted 1490 features from the dataset, reduced to less than 15% (200 most significant features) and applied three different classification approaches – Support Vector Machine, Boosting algorithms and Artificial Neural Networks.

I. Introduction

Emotion is a complex behavioral phenomenon which includes different levels of neural activations and chemical reactions in the human brain [1]. Emotion is a combination of human thought, feeling and behavior, and can be defined as a physiological reaction to different external stimuli [2]. For decades, emotions and emotion recognition have attracted a lot of attention which resulted in a variety of approaches that could be grouped into two distinct categories [2]. First group consists of methods based on non-physiological data such as speech [3] and facial expressions [4]. The advantage of this approach is the fact that the data is easily collected, without the need for any specialized and costly equipment. However, non-physiological signals can be willingly controlled which means that the individual can mask their emotion, and cause uncertainty in the classification that cannot be detected and removed. Second group relies on physiological data such as electroencephalography (EEG) [2], electromyography (EMG) [5], electrocardiography (ECG) [6], galvanic skin response (GSR) [7] etc. This approach allows better correlation with actual emotional state, but at the same time makes it harder to set up the experiment, requires special equipment and subject preparation. Noise inherently present in these signals can also present an obstacle for reliable emotion recognition.