Automatic Pain Assessment Grading Based On Facial Expressions

Pain is a complex physiological phenomenon,which involves the sensory and emotional experience of the human body,and is also a self-protection mechanism of the human body.In clinical Settings,pain intensity estimation is crucial for the diagnosis and treatment of patients,especially for patients who cannot verbally express pain.Self-report and observation methods are commonly used as the mainstream methods of pain assessment in clinical practice,but these two methods have the disadvantages of subjectivity and unfriendliness to special groups of patients.Related studies have shown that human facial expressions contain rich pain features,which makes it an effective way to extract facial features for automatic pain assessment.Based on this,this paper proposes a pain assessment model based on facial expression static images and dynamic video images.This paper is mainly divided into the following two works:(1)Based on clinicians’ experience in pain assessment and facial coding theory,we propose a graded assessment model of facial expression pain based on the fusion of global and local attention features.The model includes two modules,the global attention network and the local attention network.The local attention network is responsible for extracting representative local fine features of the face,while the global attention network extracts pain features from the whole face,so as to make up for the ignored relevant information among local fine features.Finally,global features and local features were integrated to realize the pain grading evaluation of static facial images.The test four classification accuracy of UNBC-McMaster shoulder pain dataset reaches 56.46%,which is better than other models.(2)The pain assessment model based on static images lacks the timing information between adjacent frames,which may lead to poor performance of the model.Therefore,a Dynamic Fusion Non-local Block(DFNB)based pain assessment model for dynamic video images is proposed in this paper.The model extracts the spatial features of each frame through the pre-trained Resnet101 network,and then inputs the proposed features into the Long Short-term Memory network to further extract the timing information between adjacent frames,and finally realizes the automatic evaluation of pain.The experiment shows that the DFNB module can improve the performance of pain assessment by dynamic fusion features extracted from the fourth and fifth blocks of pre-trained Resnet101 network.The accuracy of this model was86.13% on the UNBC-McMaster data set of shoulder pain dataset.In conclusion,this paper proposes a pain assessment method based on facial static/dynamic video images,which can provide an objective pain assessment auxiliary means for clinical practice,and has potential application value for clinical pain assessment of infants and aphasia and health monitoring of the elderly in nursing institutions.