Abstract:
Human Emotion Recognition has enabled state of art applications in medical healthcare,
surveillance, digital entertainment and various other sectors. Therefore, their prediction remained
an interesting aspect in the field of research. It helps in making decisions and in effective
communication. Emotions are reflected through numerous ways which includes speech
processing, behavioral changes, facial expression and physiological signals. Most of these
methods are biased. Limitations may appear in emotion recognition model because of limited
number of facial expression, fake emotions or among people with disease where they cannot
communicate their emotions. Physiological signals give a better insight of emotion classification.
This research presents a self-collection of physiological signals database for emotion
classification into 7 classes for 27 subjects using heart and skin signals. We designed a
generalized deep learning classification system for emotion detection using ECG and GSR
signals to divide signals into 7 different classes. The classification model has been built around
extracting several features and employing three different CNN architectures individually:
Alexnet, Resnet and Inception. Inception performs best among all other architectures. The
training dataset containing 5000 samples each has been used to train all three classifiers. The
proposed methodology performs reasonably well for most of the classes achieving around an
accuracy of 80% and 79.2% for ECG and GSR signals respectively. This setup is useful for the
health monitoring system and also for the investigation purpose where one can easily predict the
victim’s emotion.
