Feature Dimension Reduction For Neonatal Pain Facial Expression Recognition

At present, the methods used to assess neonatal pain are conducted artificially by professionals,and has many limitations. Using the artificial methods, plenty of time and energy are spent tocultivate professionals, and the results of the assessment are vulnerable to the subjective factors;sometimes, professionals can not assess the pain of neonates immediately. Therefore, it issignificant to develop an auto and efficient system to recognize neonatal pain, which has a widerange of applications and potential market use.The sparse representation for facial expression recognition has been one of efficient approaches.This thesis analysis the principles and processes for neonatal pain facial expression recognitionbased on sparse representation. And focus on the research of dimension reduction in neonatal painfacial expression recognition. The main contributions of this thesis are summarized as follows:(1) This thesis has researched the method of dimension reduction for neonatal facial expression.To deal with the problem that the transform vectors obtained by2DPCA are correlated, a methodcalled feature uncorrelated Two-dimensional Principal Component Analysis is proposed.(2) Applied the proposed method and other common feature dimension reduction methods arecompared in experiments of neonatal pain recognition, researched the influence of recognitionresult which is caused by the training samples number, the method of dimension reduction and thedimension of feature space. Experimental results have indicated the new method is effectiveperformance; compared with the traditional2DPCA, the new method can improve the holisticrecognition rate, and has better ability to distinguish pain and non-pain facial expression.(3)This thesis has processed the gray value of the samples in the neonatal facial expressiondatabase, obtained two sets neonatal facial expression sample which are insufficient light.Combined the new method with PCA to reduce dimension for these two sets sample, using thealgorithm framework to classify and identify these two sets sample(including four types expression),the highest recognition rate are89.33%and85.83%respectively.