Better face recognition exposes limits of image quality scores

Johns Hopkins research shows improved algorithms recover identity from degraded images, but only when useful information remains. NIST, DHS and UNHCR findings confirm low image-quality scores do not necessarily mean low recognition utility.

• Faces below roughly 8x8 pixels contain too little texture or shape for reliable matching
• NIST: removing the lowest-quality 1% of images does not eliminate false non-matches
• DHS: commercial matchers recognized many images OFIQ rated as very low quality
• UNHCR: OFIQ scores and match scores correlated only weakly to moderately across 32M images