Preservation of biometrics with AI image compression less than meets the eye

Compression of digital images can cause a loss of detail that removes the value of images containing biometric data, prompting a team of researchers to explore the potential of new compression techniques based on artificial intelligence.

The loss of data caused by strong JPEG compression can be observed visually, say the researchers from Friedrich-Alexander University Erlangen-Nurnberg in Germany, who wrote the paper: “Trustworthy Compression? Impact of AI-based Codecs on Biometrics for Law Enforcement.”

Their research indicates that AI-based compression impacts different modalities to varying degrees. Iris biometrics performance is significantly degraded with the use of such images, while fingerprint biometrics performance is barely affected.

Compression with AI codecs appears to maintain a large amount of image detail, but the researchers found that those details are not necessarily useful for biometric comparison. “Somewhat surprisingly,” they write, “in some cases they even carry less structure than JPEG images.”

The practical result is that while AI codecs may appear to preserve the level of detail needed for iris biometrics matching to the level required by law enforcement, “this appearance is misleading.”

The paper also considers “soft biometrics” like tattoos and the fabric patterns of clothing, finding that coarser patterns, like those found in polyester, carry detail through compression better, though imperfectly.

Facial recognition was not investigated for the paper, but academic research published in 2022 suggests that lossy JPEG compression is a barrier to face biometrics performance, and may contribute to system bias.

While there is an established standard for fingerprint image quality, which is currently being refined, the standard and associated tool for assessing image quality for face biometrics is yet to reach final approval.

Article Topics

AI  |  biometric data quality  |  biometrics  |  biometrics research  |  forensics  |  image compression  |  research and development