Centroid-Based Representation for Gender Classification Using Celebrity Cartoon Images

Abstract

This work focuses on the exploitation of the notion of representing celebrity cartoon genders by clustering based on the cluster centroid. In this approach, features and samples are systematically reduced to provide a compact representation. For the extraction of deep features, the FaceNet architecture is used. The K-means algorithm is used to cluster celebrity cartoons based on their gender. The approach is carried forward to preserve the cluster representatives of each cluster for further classification and sample space reduction. We also explored suitable conventional classifiers and recommended the K-Nearest Neighbor (KNN) classifier, which is suitable for centroid-based gender classification. Further, the well-known subspace techniques such as Principal Component Analysis (PCA) and Fisher Linear Discriminant (FLD) have been adopted for dimensionality reduction. To demonstrate the effectiveness of the proposed model, we conducted extensive experiments on a dataset specific to celebrity cartoon images, namely IIIT-CFW. In contrast to existing methods, our model achieved a state-of-the-art result with an F-measure of 92.99%.