Manifold Learning Algorithm For Facial Expression Recognition Of Pain In Neonates

The traditional method, used to assess neonatal pain and conducted by health professionals, has many limitations. For example, using this artificial method, plenty of time and energy are spent to cultivate many health professionals; the results of the assessment are vulnerable to the subjective factors; Sometimes, health professionals cannot assess the neonatal pain immediately at the hospital because of some reasons. With the development of technologies of face recognition and facial expression recognition, using computer to auto assess the neonatal pain is becoming possible now. Therefore, it is significant to develop an auto and efficient system to recognize neonatal pain.Feature extraction is the key technology in the procedure of neonatal pain recognition and it determines the recognition rate. So, the aim of the thesis is to find an efficient and robust feature extraction algorithm to correctly recognize the neonatal pain. Many researches have shown that face images distribute in an un-linear manifold space, and the recently proposed manifold learning method is very efficient to find the un-linear structure embedding in the data. By deeply investigating the manifold learning methods, this thesis proposes two manifold learning algorithms and the main contributions of this thesis are summarized as follows:1) A comprehensive overview about the current situation of facial expression recognition and neonatal pain recognition are given. Deeply discussion is made on some classical manifold learning methods as the theoretic supporting of the thesis's research.2) An eastern neonatal expression database is established as the research needed.3) To deal with the problem that the transform vectors obtained by linear manifold learning methods are often non-orthogonal, a new method called Orthogonal Isomatric Projection (O-IsoProjection) is proposed. O-IsoProjection not only provides an optimal approximation to the true isometric embedding of the underlying data manifold and gives a more faithful representation of the data's global structure, but also can preserve the metric structure of face subspace。4) In the area of pattern recognition, a general rule for feature extraction is to extract features as uncorrelated as possible, but the discriminant features in linear manifold learning algorithms are correlated. In order to obtain an optimal set of uncorrelated discriminant features, Uncorrelated Locality Sensitive Discriminant Analysis (U-LSDA) is presented. With the restriction of the uncorrelated, facial expression features obtained by the new method are uncorrelated, and the uncorrelated characteristic increases the discriminant of the features extracted by U-LSDA.5) The proposed methods and other manifold learning methods are compared in experiments of neonatal pain recognition, and experimental results have indicated its effective performance.