Lossy image compression negatively affects facial recognition algorithms, research shows

Experimental research by a team from the Department of Computer Science in the University of Durham, UK, demonstrates that image compression techniques that remove data have a negative bearing on the way facial recognition algorithms function with regard to data training and testing. The affects may contribute to biometric bias, the researchers say.

Lossy compression is the process by which some data is removed from an image file in order to reduce its size or other original properties. The data removed from such a file using lossy techniques cannot be restored.

The study carried out by a quartet of academics investigates the impact of lossy JPEG compression algorithm on contemporary facial recognition performance, according to an abstract of the research findings.

The researchers say there is a gap in how this impact varies with different race groups. Their experiment finds that common compression methods can impact facial recognition performance by up to 34.55 percent for racial phenotype categories like darker skin tones.

Further, they found that removing chroma subsampling (a type of compression that reduces color information in the image) during compression improves a system’s false match rate by up to 15.95 percent across all affected groups, “including darker skin tones, wide noses, big lips, and monolid eye categories,” the abstract explains.

Overall, the evaluation finds that using lossy compressed facial image
samples for matching decreases performance more significantly on specific phenotypes.

However, the use of compressed imagery during training does make the resulting models more resilient and limits the performance degradation encountered, as
lower performance amongst specific racially-aligned subgroups remains, notes the report in its conclusion.

The paper also considers the impact of factors such as balanced and unbalanced training datasets and compression levels.

The authors add that their work looking at the impact of lossy compression algorithms on phenotype-based racial groups is part of stakeholder efforts in making available additional evidence-based insights and understanding to guide the mitigation of bias in the development of future face biometrics algorithms and systems.

NIST’s Patrick Grother noted the benefit of low image compression to address demographic issues in a presentation in late-2020.

Other research efforts published recently have tried to find better ways to quantify demographic disparities in the effectiveness of facial recognition systems.

Article Topics

accuracy  |  biometric bias  |  biometrics  |  biometrics research  |  facial recognition  |  image compression