光学学报, 2008, 28 (11): 2109, 网络出版: 2008-11-17
新生儿疼痛面部表情识别方法的研究
Research on Recognition for Facial Expression of Pain in Neonates
图像处理 表情识别 支持向量机 新生儿疼痛 AdaBoost算法 Gabor变换 image processing expression recognition support vector machine neonatal pain AdaBoost algorithm Gabor transform
摘要
针对新生儿的疼痛与非疼痛面部表情识别, 提出将Gabor变换和支持向量机(SVM)相结合的分类识别方法。对归一化后的大小为112 pixel×92 pixel的新生儿面部图像进行二维Gabor小波变换, 提取出 412160维Gabor特征; 针对Gabor 特征向量维数高、冗余大的特点, 采用Adaboost算法作为特征选择工具, 去除冗余的Gabor特征, 从412160 维特征中选取出900维Gabor特征; 对选取出的Gabor特征用SVM进行疼痛表情的分类识别。该方法综合运用Gabor 特征对于面部表情的良好表征能力、AdaBoost 算法的特征选择能力以及SVM 在处理少样本、高维数问题中的优势。对510幅新生儿的表情图像进行测试的结果表明, 疼痛与非疼痛表情的分类识别率达到85.29%, 疼痛与安静表情的分类识别率达到94.24%, 疼痛与哭表情的分类识别率达到78.24%。
Abstract
A classification method to distinguish the neonatal pain expression from non-pain expression is proposed, which combines Gabor wavelet transform with support vector machine (SVM). At first, each neonatal facial image, which is normalized to the size of 112 pixel×92 pixel, is transformed by the 2D Gabor wavelet to extract 412160 Gabor features. Since the high-dimensional Gabor feature vectors are quite redundant, AdaBoost is introduced as a feature selection tool to remove the redundant ones. In experiments, 900 features are selected from 412160 original Gabor features. Finally, the selected Gabor features are fed into the SVM for final classification. This method takes the advantages of the favorable ability of Gabor feature in representing facial expression, the effective function of Adaboost in feature selection, and the high performance of SVM in the solution to small sample size, high dimension problems. Experiments with 510 neonatal expression images show that the method is quite effective. The best recognition rates of pain versus non-pain (85.29%), pain versus calm (94.24%), pain versus cry (78.24%) are obtained.
卢官明, 李晓南, 李海波. 新生儿疼痛面部表情识别方法的研究[J]. 光学学报, 2008, 28(11): 2109. Lu Guanming, Li Xiaonan, Li Haibo. Research on Recognition for Facial Expression of Pain in Neonates[J]. Acta Optica Sinica, 2008, 28(11): 2109.