An Ensemble of Deep Convolutional Neural Networks for Drunkenness Detection Using Thermal Infrared Facial Imagery

This paper proposes an original method for subject independent drunkenness detection using an ensemble of Deep Convolutional Neural Networks (DCNNs) for processing of thermal infrared facial imagery characterizing the subjects to be tested. The proposed neural system consists of an ensemble of two DCNNs modules for thermal infrared facial image processing; the first module is composed by 12 layers and the second one has 10 layers. The two DCNNs have been trained separately, using different architectures and different sets of parameters. The final decision is influenced by the confidence degrees of two CNN component modules. The proposed method is evaluated using the dataset of 400 thermal infrared facial images belonging to 10 subjects. For each subject the dataset contains 20 thermal images corresponding to sober condition and other 20 images for inebriation condition obtained 30 minutes after the subject has drunk 100 ml amount of whisky. The experiments of the proposed DCNN couple for subject independent drunkenness detection lead to the overall correct detection score of 95.75%. This confirms the effectiveness of the proposed approach.