Stress Detection through Thermal Facial Images Inspired by Characteristics of Psychophysiological Signals

Stress detection based on physiological signals requires numerous contact sensors that can cause anxiety and disturbance for the subject and may lead to invalid results. Using thermal imaging in the stress detection has several advantages including being impalpable, non-contact and fast installation. Many researches have used thermal images in deception detection. Most of the previous articles used simple features such as average and increasing temperature slopes for analysis. This study uses the features defined in the earlier references as well as defining a set of features based on the principle of solidarity between temperature changes in different areas of the face with other physiological signals during stress interactions.
A database including 13 subjects through a lie detection protocol while physiological signals are recorded as well as thermal images is provided in this study. The proposed features are applied on the database. Independent component analysis and linear discriminate analysis is used for feature reduction and classification, respectively. In addition, a manual scoring method according to manual scoring methods for physiological signals is introduced for manual scoring thermal images.
The accuracy of stress detection in each trial (a question) based on the proposed features becomes more than the ones using only features defined in previous studies (an increase of 9.44% in test data and 27.33% in training data). The accuracy of detecting the target question (the question which the subject has lied) is 80%. Two manual scoring methods, thermal and physiological have the correlation of 70%.
According to the results, using the features which is usually defined on physiological signals including respiratory, galvanic skin response, heart rate and blood pressure lead to improvement in stress detection based on thermal images. Therefore, the efficiency of the proposed method based on defined features and proper classification for the detection of deception are confirmed using thermal images.