LipVerify is based upon several years of research undertaken within the School of Electronics and Electrical Engineering and Computer Science at Queen’s University Belfast. A set of rigorous studies, including several dedicated PhD programmes, have identified the potential strength of lip-movement (viseme) biometrics for authentication. Several research papers emanating from these PhDs have been rated as “internationally excellent in terms of originality, significance and rigour”.
The LipVerify Biometric matching process uses a sequence of image and video processing techniques to accurately detect and track the user’s mouth region. Further novel image processing techniques are then used to extract highly discriminative speaker-specific features representing the lip appearance and dynamics, taking care to remove the effects of pose and illumination variation.
During enrolment to the system, state-of-the-art Machine Learning processes are employed to build models of each user’s visual speech behaviour, known as their ‘viseme profile’, and these models are used during verification to estimate the likelihood that the claimed person produced the video sequence which was provided.
Additionally significant work has been undertaken in the following areas to ensure the technology is sufficiently robust for commercial deployment –
- Illumination Compensation – algorithmic improvements to help eradicate effects of both global (e.g. sunlight, room lighting) and local (e.g. nose shadow) lighting effects.
- Automated Lip Alignment –algorithmic improvements to automatically ‘find’ the user’s lip area. This improve’s front-end usability of the solution and overall accuracy.
- Optimise Enrolment/Verification Data Requirements – refinement and optimisation of the volume/content of enrolment and verification data to ensure strong verification accuracy and excellent end-user experience.
The techniques outlined above, coupled with embedded Visual Phrase Recognition (VPR), ensure that LipVerify is robust, easy-to-use and highly resistant to compromise.