Checking photo ID is necessary for verifying identity, enabling people to cross borders, buy age-restricted goods and access services. In a similar way, CCTV identification relies on matching footage to mugshots of suspects. In both situations, the to-be-matched images feature people the viewers are unlikely to know personally. The error rate when attempting to decide whether two images feature the same unfamiliar person is high, even among experienced police and other security practitioners. This is problematic because incorrect decisions can have serious consequences, undermining forensic evidence or even national security.
Our ongoing work focuses on reducing incorrect identification decisions, leading to positive benefits in forensic and security contexts. We have developed a novel viewing procedure that enables the user to ‘interact’ with a face they are trying to match by clicking and dragging the image into any orientation, revealing how the person might look from different viewpoints. The procedure provides the user with knowledge of the 3D facial structure of the person shown in the to-be-matched image. If you have never encountered a person before this knowledge is helpful, and our experiments show that it decreases error when attempting to match static faces. The innovation is low cost and easy to implement: Interactive faces can be easily rendered from videos of faces moving from side to side. We have designed the system with law enforcement stakeholders so that it can be integrated into existing police systems.
As part of our research, which was recently published in the British Journal of Psychology (https://doi.org/10.1111/bjop.12499), we tested over 650 participants. We tested people who have typical face recognition skills, as well as those who have exceptional face recognition skills, known as superior recognizers. While typical recognizers are likely to be recruited to identity verification roles, superior face recognizers are extremely sought after in security contexts. An innovation that improves the performance of both typical and superior recognizers has significant security implications.
In each experiment the participants were asked to decide whether two images featured the same person. One of these images was a Facebook profile picture, and the to-be-matched image was either a static image of a face or an interactive image of a face.
Our results were consistent across four experiments, revealing that the interactive procedure reliably improves face-matching accuracy for both typical and superior face recognizers. The procedure also improves accuracy when image quality is low, for example as it might be in CCTV footage: accuracy was higher using the interactive procedure, even when the Facebook profile image was pixelated. These results are significant. This is the first time that research has revealed a way of optimizing superior face recognizer performance to minimise the risk of erroneous face matching decisions in practice.
Our ongoing research seeks to thoroughly explore and account for the interactive face matching advantage in order to further improve face matching performance and develop training for end users.
