Pose-Invariant Face Detection by Replacing Deep Neurons with Capsules for Thermal Imagery in Telemedicine

The aim of this work was to examine the potential of thermal imaging as a cost-effective tool for convenient, nonintrusive remote monitoring of elderly people in different possible head orientations, without imposing specific behavior on users, e.g., looking toward the camera. Illumination and pose invariant head tracking is important for many medical applications as it can provide information, e.g., about vital signs, sensory experiences, injuries, wellbeing. In the performed experiments, we investigated the influence of different modifications of images (rotation, displacement of facial features, and displacement of facial quarters) on the prediction accuracy. Specifically, two models were tested on the set of collected low-resolution thermal images: Inception V3 Convolutional Neural Network (CNN) and Hinton's Capsule Network. The preliminary results confirm that the prediction ability of the model based on capsules can deal with different head orientations much better than CNN (for the 45° head rotation Capsule Network achieved ~% accuracy while CNN only 9.5%).