Abstract

Caricatures can help to understand the perception of a face. The prominent facial feature of a subject can be exaggerated, so the subject can be easily identified by humans. Recently, significant progress has been made to face detection and recognition from images. However, the matching of caricature with photographs is a difficult task. This is due to exaggerated features, representation of modalities, and different styles adopted by artists. This study proposed a cross-domain qualitative feature-based approach to match caricature with a mugshot. The proposed approach uses Haar-like features for the detection of the face and other facial attributes. A point distribution measure is used to locate the exaggerated features. Furthermore, the ratio between different facial features was computed using different vertical and horizontal distances. These ratios were used to calculate the difference vector which is used as input to different machine and deep learning models. In order to attain better performance, stratified k-fold cross-validation with hyperparameter tuning is used. Convolution neural network-based implementation outperformed the machine learning-based models.