Classification of Gender in Celebrity Cartoon Images

Abstract

This paper presents a model by integrating deep architecture-based feature extraction with classical learning algorithms for the effective gender classification of celebrity cartoon images. The proposed model makes use of colored celebrity cartoon images. For face representation, we have extracted features using various conventional approaches such as Scale-Invariant Feature Transform (SIFT), Fast Fourier Transform (FFT), and Discrete Cosine Transform (DCT), and the deep approach using deep features from a pre-trained architecture. After extracting the features to classify the gender of celebrity cartoon images, we have used conventional classifiers. Here in this work, we have recommended the integration of random forest (RF) with deep features extracted from FaceNet pre-trained architecture. We have used an existing dataset, the Cartoon Faces in the wild (IIIT-CFW) database of celebrity cartoon faces consisting of 8928 cartoon images. At the same time, we segregate it into two distinct classes, male and female, comprising 4594 and 2073, respectively. There is an imbalance in the proportion of the dataset; we have attempted to balance the unbalanced data using the SMOTE technique. Lastly, to classify the gender of celebrity cartoon images, we have adopted distinct supervised conventional learning models. The classification results are validated using classification metrics such as accuracy, precision, recall, f-measure. We show that our approach model establishes a state-of-the-art F1-score of 95% on the task of gender recognition of cartoons faces. The model has been compared against other existing contemporary models for its effectiveness.