Multi-scale regional attention networks for pain estimation

Pain is a common treatment response of patients in clinical medicine, which indicates the patient's status. Of late, automatic pain estimation methods have received more and more attention due to their convenience and objectivity. However, the previous researches mostly employ existing models or frameworks, without the exploration of pain locality. In this paper, we propose Multi-Scale Regional Attention Networks (MSRAN), to adaptively capture the importance of facial pain regions. Specifically, the proposed MSRAN aggregates and embeds varied number of multi-scale region features by a convolutional neural network with self-attention module for the locality of pain. Then, the proposed relation-attention module is leveraged to explore the relationship of pain regions. Last, the well-designed loss function is employed to increase the discrimination of model. We validate our MSRAN on the public pain dataset, which shows the effectiveness of the proposed method.