Gender classification using discrete cosine transformation: a comparison of different classifiers

We have investigated the problem of gender classification using a library of four hundred standard frontal facial images employing five classifiers, namely k-means, k-nearest neighbors, linear discriminant analysis (LDA), Mahalanobis distance based (MDB) classifiers and our modified KNN classifier. The image data independent discrete cosine transformation (DCT) basis is used for facial feature extraction. Areas under the convex hull (AUCH) of the classifiers are measured by varying the values of threshold for each feature subset in the receiver operating characteristics (ROC) curve. The scalar values of AUCH of the ROC curve increases with increasing number of features. More features yield a better representation of the gender facial image. The overall performance of classifiers is compared with different values of AUCH versus features under different conditions. It has been observed that when the number of features is increased beyond 5, AUCH starts to saturate. Our experimental results demonstrate that modified-KNN performs better than the rest of the conventional classifiers under all conditions. The LDA classifier did not perform well in the DCT domain; however, it gradually improved its performance with increasing number of features