The growing importance of digital identity for accessing various services has led to the popularity of biometric methods, with facial recognition being the most preferred due to its non-intrusive nature. However, challenges like poor image quality can affect its accuracy.
T Stamp Inc (NASDAQ:IDAI, EURONEXT:AIID), the Atlanta-based global provider of identity services doing business as Trust Stamp, has developed an AI-powered suite of facial attribute and image quality assessment tools (FAIQA) to assess facial attributes and image quality, aiming to improve success rates in enrollment and re-authentication in various applications.
In a recently-released white paper, Trust Stamp’s chief science officer Norman Poh explains how the technology uses automatically estimated facial attributes and quality measures to enhance the performance of facial recognition systems.
Below are excerpts from the white paper written by Poh along with Leanne Attard and Reuben Farrugia.
Facial recognition's role in digital identity
Being able to assert one’s own identity digitally provides individuals with convenient access to numerous services spanning finance, employment, travel, legal, and healthcare. This has led to a steady increase in popularity of the use of biometric modalities. Among all biometrics, face is the most popular and advantageous because it is non-intrusive and does not require any contact capture, which is an important consideration post pandemic.
Moreover, the accuracy of face recognition has improved markedly thanks to the advancements of deep learning architectures that continue to reduce in size and can benefit from even larger training data sets. Unfortunately, the output of such systems can still be adversely affected by poor image quality due to e.g., suboptimal lighting conditions, an inadequate head pose or accessories that the users wear. In this article, automatically estimated facial attributes and derived face-related quality measures are explored to improve the usability and performance of automatic face recognition solutions.
Navigating facial image quality in diverse settings
Due to the improvement in the quality and availability of capture devices such as smartphones, tablets and webcams, authentication applications are increasingly available to the general public. In uncontrolled environments, this can sometimes produce low quality images due to poor illumination, movement, non-frontal pose, facial occlusions, and other factors. Given these challenges, the ability to analyze the quality of a facial image is essential for any face processing task, including face recognition, presentation attack detection (PAD) and age estimation amongst others.
Trust Stamp's AI-powered solution
To address issues related to unfavorable facial capture, Trust Stamp has built an AI-powered suite of facial attribute and image quality assessment tools (FAIQA) which can be integrated into any such applications with the objective of improving the quality of the capture and thereby improving success in both enrollments and re-authentications.
Find out more information about FAIQA here.
Use cases
There are several scenarios where a biometric sample quality’s assessment can be used. Use cases for FAIQA include, but are not limited to:
- Image rejection threshold: Images that do not satisfy one or multiple quality metrics can be rejected. This ensures that only high-quality biometric samples are captured, thereby improving the facial image recognition accuracy.
- Acquisition feedback: Using individual quality metrics and attributes, specific feedback can be given to the user during a guided capture session, resulting in an improved user experience in two ways: lowering the friction by reducing the number of retakes and improving the likelihood of a successful recognition thanks to the improved image quality.
- Quality summarization: Quality can also be monitored by recording it over time for different capture devices, locations, and subjects. This generates log files that allow Trust Stamp to identify problematic devices, difficult locations, and subjects who consistently submit low quality samples so that appropriate interventions can be made.
- Video frame selection: Images in a video sequence used for face recognition and proof of liveness can be ranked and selected by their assigned quality scores to improve the application’s performance and confidence.
- Database maintenance: Existing images in a database can be ranked and filtered by quality such that biometric matching against better quality images is favored to ensure a higher identity matching accuracy.