The Forty-Pixel Face
An underground parking garage camera records 320×240 at 10 fps with strong noise in low light. A suspect stands 8 m from the camera; the distance between the eyes is about 40 pixels in the best frame, and the face is turned 20° from frontal. An investigator brings you a still that a colleague processed with a deconvolution filter and an AI upscaler; it now shows crisp eyebrows, a scar, and a recognizable likeness to a named person. The investigator asks you to sign a report stating the image is 'forensically enhanced' and supports identification. You have the original clip and 14 adjacent frames. The camera's compression is H.264 with a 2-second GOP and visible blocking.
1. Should you sign the report as requested? Explain your reasoning.
Read the worked answer
No. At 40 pixels between the eyes, the face occupies a tiny fraction of a 320×240 frame; the fine detail needed for a scar or a confident likeness is not present in the original. Deconvolution can partially restore edges if the point-spread function is known, but an AI upscaler is generative: it predicts plausible high-frequency texture from training data and can invent features. A scar that appears only after upscaling is not evidence. Signing 'forensically enhanced' would overstate reliability and conflate restoration with synthesis. The correct response is to deliver the original and the processed image as separate, labeled products, state the processing chain and its limitations, and decline an identification opinion. The investigator may use the processed image as an investigative lead, not as a basis for identification.
2. How would you validate whether the enhancement reveals real detail or invents it?
Read the worked answer
Run the same pipeline on a known reference in the scene-a sign, a license plate, or a measured edge-and see whether it preserves known detail or adds plausible but false structure. Compare all 14 adjacent frames: real facial features should remain stable across frames with small pose changes, while hallucinated features flicker or change. Check whether the scar exists in any unprocessed frame at the correct location and whether its shape is consistent with the illumination. Try a second, different upscaler; disagreement indicates instability. Estimate effective resolution: 40 pixels between the eyes may allow a few pixels per feature, not enough for a scar. Document each test. If the feature cannot be demonstrated in the original data, the enhancement cannot be used as identification evidence.
3. Draft the conclusion and limitation language for your report.
Read the worked answer
Write: 'No identification opinion is offered. The original image provides approximately 40 pixels between the eyes, with a 20° pose and H.264 compression; this is insufficient for a reliable facial identification. The enhanced image is a processed derivative and may contain synthetic detail introduced by deconvolution or upscaling. It is suitable only as an investigative lead.' Then describe the processing chain, software versions, settings, and the validation tests performed. Include the original frame and the processed image side by side with clear labels. State that the processed image must not be presented as if it were the original, and that any comparison to a named person should be made from the original data by a qualified examiner using documented methods.