Neonatal Pain Facial Expression Recognition Based On Orthogonal Matching Pursuit Algorithm

Nowadays, medical research has proven that repeated pain could bring a negative impact toinfants. Studies about neonatal pain have been concerned by lots of researchers. Assessment tools ofneonatal pain have been much needed because neonates couldn’t tell their subjective feelings.Among all the assessment indicators, the infants’ facial expression is considered as the most reliableand effective one. If the recognition work was completed by the well trained health professors, it’llbe a time costly and difficult job, meanwhile, the recognition results are vulnerable influenced bythe subjective factors. So the research about facial expression recognition is very useful andmeaningful, and a kind of auto neonatal pain assessment system is much needed.As a new theory of signal processing, compressive sensing theory has plenty of advantages. It ismainly used under the condition which the signal is compressive or sparse. It can compress a signalwhen the signal is sampled. This theory is possible used in the area of pattern recognition and imageprocessing, because image signals are always sparse. The research work in this paper mainlyincludes the followed aspects.(1)We attract features of the neonatal facial expression we used in this paper withdown-sampled and primary component analysis separately.(2)We did research on a new kind of neonatal pain facial expression recognition method, whichis based on orthogonal matching pursuit algorithm. With this method, the infant facial expression iseasy to recognize and classify because the sparse representation of the test infant facial expressionimage processed by the orthogonal matching pursuit algorithm has distinct class information.(3)The research in this paper proved that the method we raised can achieve a good recognitionrate. The experiment results show that when the number of the training sample and abstractedfeatures are sufficient, and the sparse representation is correctly found, the algorithm frameworkpresented in this paper can work well with the feature abstraction such as down-sampled or primarycomponent analysis.