Effects of Training Set Dimension on Recognition of Dysmorphic Faces with Statistical Classifiers

Şafak Saraydemir¹, Necmi Taşpınar², Osman Eroğul³, and Hülya Kayserili⁴

¹Department of Electronics Engineering, Turkish Military Academy, Turkey

²Department of Electrical and Electronics Engineering, Erciyes University, Turkey

³Biomedical Engineering Centre, Gülhane Military Medicine Academy, Turkey

⁴Department of Medical Genetics, İstanbul University Medicine Faculty, Turkey

 Abstract: In this paper, an evaluation using various training data sets for discrimination of dysmorphic facial features with distinctive information will be presented. We utilize Gabor Wavelet Transform (GWT) as feature extractor, K-Nearest Neighbor (KNN) and Support Vector Machines (SVM) as statistical classifiers. We analyzed the classification accuracy according to increasing dimension of training data set, selecting kernel function for SVM and distance metric for kNN. At the end of the overall classification task, GWT - SVM approach with Radial Basis Function (RBF) kernel type achieved the best classification accuracy rate as 97.5% with 400 images in training data set.

 Keywords: Dysmorphology, GWT, principal component analysis, face recognition, SVM, KNN.

Received February 26, 2013; accepted May 6, 2013

Full Text