EEG-Based Emotion Estimation with Different Deep Learning Models

Abstract

Emotion has a vital role in people's routine lives. It can be expressed via voice, facial expressions, body languages, mimics with intentionally or unintentionally to interact with the environment. In this regard, it is required to understand the emotion better to interpret the emotions. Emotion is generally used in many areas including rehabilitation applications, braincomputer interactions, genome-wide applications, healthcare services etc. There are many studies exist about emotion recognition with different approaches based on facial expression, voice and physiological signals. Yet, the first two of them can give incorrect information about emotions since these approaches can be manipulated by subjects easily. Thus, the more reliable and more durable approach proposed including EEG signals. Although it gives valuable information on emotion, EEG-based emotion estimation applications have not reached the desired level since its abstract and pattern recognition methods (falsified feature extraction methods, false classifier algorithms, big data, etc.) used for that applications. EEG-based emotion estimation is a complicated assignment which requires deep features, many EEG channels, clear signals and classifier algorithms. Determining the features and analyzing them requires time, thus in this study, we applied deep learning to discriminate the positive/negative emotional states. Our proposed method includes three parts; i) Collecting EEG data ii) Preprocessed the EEG data to denoise the signal iii) Deep learning with AlexNet and VGG-16 We collected EEG signals from 28 various subjects aged between 21-28 via portable and wearable EEG device called Emotiv Epoc+ 14 channel. In order to collect the signals, we applied four different video games as stimuli (2 negative and 2 positive labelled games) and collected signals totally 20 minutes long for each subject. At the end of the EEG collection process, we obtained 1568 number of EEG samples (14x28x4). To collect more reliable and healthy information from signals we preprocessed our signals. Finally, we performed two different deep learning algorithms to determine the positive-negative emotions and to compare their results. It is observed that the classification accuracies differ with different algorithms and the classification performance was found 92,09% with VGG16 which is superior to AlexNet algorithm 87,76%.