Automatic pain recognition from facial expressions is a challenging problem that has attracted a significant attention from the research community. This article provides a comprehensive analysis on the topic by comparing some popular and Off-the-Shell CNN (Convolutional Neural Network) architectures, including MobileNet, GoogleNet, ResNeXt-50, ResNet18, and DenseNet-161. We use these networks in two distinct modes: stand alone mode or feature extractor mode. In stand alone mode, the models (i.e., the networks) are used for directly estimating the pain. In feature extractor mode, the “values” of the middle layers are extracted and used as inputs to classifiers, such as SVR (Support Vector Regression) and RFR (Random Forest Regression). We perform extensive experiments on the benchmarking and publicly available database called UNBC-McMaster Shoulder Pain. The obtained results are interesting as they give valuable insights into the usefulness of the hidden CNN layers for automatic pain estimation.

Pain assessment is a challenging problem in the field of emotion recognition. Pain represents a complex emotion difficult to detect or to estimate its intensity. This is what makes automatic pain assessment playing an important role in clinical diagnosis. Taking into consideration that pain generally generates spontaneous facial behaviour, these facial expressions could be used to detect the presence of pain. As a matter of fact, previous researches used machine learning and deep learning either to detect pain or to estimate pain level. In this paper, we propose a fine-tuning of pre-trained data-efficient image transformers and distillation (Deit) for pain detection from facial expressions. The effectiveness of the proposed architecture is evaluated on two publicly available databases, namely UNBC McMaster Shoulder Pain and BioVid Heat Pain. The proposed approach achieved promising preliminary results compared to the state of the art.