Hybrid learning-based model for exaggeration style of facial caricature

• Main Author: Sadimon, Suriati
• Format: Thesis
• Language: English
• Published: 2017
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.utm.my/id/eprint/78996/1/SuriatiSadimonPFC2017.pdf; http://eprints.utm.my/id/eprint/78996/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:107415
• Institution: Universiti Teknologi Malaysia

Prediction of facial caricature based on exaggeration style of a particular artist is a significant task in computer generated caricature in order to produce an artistic facial caricature that is very similar to the real artist’s work without the need for skilled user (artist) input. The exaggeration style of an artist is difficult to be coded in algorithmic method. Fortunately, artificial neural network, which possesses self-learning and generalization ability, has shown great promise in addressing the problem of capturing and learning an artist’s style to predict a facial caricature. However, one of the main issues faced by this study is inconsistent artist style due to human factors and limited collection on image-caricature pair data. Thus, this study proposes facial caricature dataset preparation process to get good quality dataset which captures the artist’s exaggeration style and a hybrid model to generalize the inconsistent style so that a better, more accurate prediction can be obtained even using small amount of dataset. The proposed data preparation process involves facial features parameter extraction based on landmark-based geometric morphometric and modified data normalization method based on Procrustes superimposition method. The proposed hybrid model (BP-GANN) combines Backpropagation Neural Network (BPNN) and Genetic Algorithm Neural Network (GANN). The experimental result shows that the proposed hybrid BP-GANN model is outperform the traditional hybrid GA-BPNN model, individual BPNN model and individual GANN model. The modified Procrustes superimposition method also produces a better quality dataset than the original one.