Facial Expression Synthesis using Kernel Approach

• Main Author: Marcella, Peter
• Format: Thesis
• Language: English
• Published: Universiti Malaysia Sarawak (UNIMAS) 2020
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/32957/3/Facial%20Expression%20Synthesis%20using%20Kernel%20Approach.pdf; http://ir.unimas.my/id/eprint/32957/
• Institution: Universiti Malaysia Sarawak

Recently, facial identity and emotion study has gained some interest from researchers especially in the works of integrating human emotions and machine learning to improve the current lifestyle. Emotions are initially expressed through facial expression and followed by body language to deliver information. By nature, emotions can be easily expressed such as happiness, sadness, and surprised. However, in computer language, it is still a challenging task especially synthesising realistic facial expression. Therefore, various methods have been proposed to synthesise better facial expression systems that include learning-based and statistical-based approaches. Most of these approaches applied linear methods and the most commonly used one is the Principal Component Analysis (PCA). PCA is a linear transformation technique and can be used for reducing high dimensional data, extracting facial features from an input, transforming the extracted features to represent a face via a face model and subsequently extended for face recognition system. However, linear transformations may lead to some information loss along the way. Furthermore, the facial structure of a face in itself is complex to be expressed using a linear method. Therefore, in this study, a kernel-based method is proposed to deal with the linear approach problems on transformation and projection. This study explored the potential of using nonlinear kernel approach for synthesising neutral facial expressions 3D geometric face models to improve the performance and recognition rates. The kernel approach employed in the research is a novel modified kernel-based Active Shape Model whereby it employed mean template-based face model. The results from the modified kernel method is then compared with the linear-based Active Shape Model and the outcome of the face recognition is used to evaluate the resulting synthesised neutral facial expression. Experiment results have recorded the highest recognition rate with 100% of true positive and have also shown that the recognition iv outperformed the linear Active Shape Model. The qualitative results of the synthesis have also shown the almost (if not) real facial expressions of the subject. In conclusion, the proposed modified kernel-based Active Shape Model using template-based approach can improve the synthesis of facial expression which then would increase the performance of the recognition rates. Future work would include to further investigating the effect of adjusting expression intensity on the shape model of the synthesised facial expression by integrating the nonlinear approach into an automated face recognition system and applying optimisation approach to improve the efficiency of the modified kernel-based Active Shape Model.