Towards Improved Cartoon Face Detection and Recognition Systems

Given the significant advancement in face detection and recognition techniques for human faces over recent years, we questioned how well would they work for cartoon faces a domain that remains largely unexplored yet, mainly due to the unavailability of abundant data sets and failure of traditional methods on these. In the present article, we employ various state-of-the-art deep learning frameworks for detecting and recognizing faces of cartoon characters along with proposing a novel approach to cartoon face recognition. For face detection, our work demonstrates the effectiveness of the Multi-task Cascaded Convolutional Network (MTCNN) architecture [1] and contrasts it with other benchmark methods. For face recognition, we present two feature-based techniques: (i) an Inductive transfer approach combining the feature learning capability of the Inception v3 network [2] and feature recognizing capability of Support Vector Machines (SVM), (ii) a proposed Hybrid Convolutional Neural Network (HCNN) based recognition framework trained over a fusion of pixel values and 15 manually located facial keypoints. All the methods are evaluated on the Cartoon Faces in the Wild (IIIT-CFW) database. We show a detailed analysis of the performance of the models using several metrics over a number of input constraints. Experiments show that the MTCNN based model results in a respective gain of 3.97%, 1.19% and 2.78% on the True positive rate, False positive rate and False negative rate over the state-of-the-art detection method while both the recognition models surpass the state-of-the-art in terms of F-score. The Inception v3+SVM recognition model also establishes a new benchmark F-score of 0.910 on the task of cartoon gender recognition. We also introduce a small sized database containing location coordinates of 15 key points of the cartoon faces belonging to 50 public figures included in the IIIT-CFW database.