Facial Expression Recognition of Pain in Neonates

This paper proposes a pain expression recognition method using boosted Gabor features. At first, each neonatal facial image which is normalized to the size of 112times92 pixels is convoluted with the 2D Gabor filters to extract 412160 Gabor features. Since the high-dimensional Gabor feature vectors are quite redundant, we propose a modified version of AdaBoost algorithm, called the HybridBoost, to select the most informative features for classification. Then, this paper proposes a coarse-to-fine hierarchical classifier to treat the "pain vs. nonpain" problem as two sub-problems: "calm vs. noncalm" and "cry vs. pain". Experiments with 510 neonatal expression images show that the proposed method is quite effective. Only 30 Gabor features are enough to achieve a good classification performance. The recognition rate of pain vs. nonpain is up to 88%. The result indicates a high potential for developing an automatic neonatal pain assessment system.