An AI that can repair blurred or distorted images of fingerprints lifted from crime scenes could make identifying people easier, but it is unclear whether such evidence would stand up in court.
at West Virginia University and his colleagues trained an AI to cancel out distortions of fingerprints caused by incorrect camera focusing and other errors.
The team took a data set of 15,860 clean fingerprint images from 250 subjects and created blurred versions of them synthetically at varying levels of distortion. Almost 14,000 of these pairs of images were used to train an AI and the others were used to test its performance.
Advertisement
The researchers created a generative adversarial network where one neural network is pitched against another. One attempts to generate realistic deblurred fingerprints, while the other assesses them for realism, and the results are fed back to each algorithm to enable them to improve.
Because the ridges and valleys in fingerprints are key to identification, the team used a separate algorithm to highlight those features during training and check that the deblurring algorithm didn鈥檛 destroy that information or add to it erroneously.
The researchers found that their model could achieve 96 per cent accuracy at the lower end of their range of blurring intensity and 86 per cent at the higher end. Intuitively, the algorithm was more successful at extracting a clean image the less the input image was blurred.
But at Forensic Equity, a UK firm providing forensic science services, says that blurred images due to human error should be avoidable with the correct training. He also warns that the black-box nature of neural networks would present a problem in court, where any manipulation of fingerprint images would be the focus of scrutiny.
Neural networks can be trained to perform a wide range of tasks, but they aren鈥檛 auditable like human-generated code. Often, the inner workings of the resulting models are a mystery even to their creators.
Goodwin says UK courts would be unlikely to accept evidence that has been manipulated by AI unless there is an audit trail or an explanation of what processing has been done.
鈥淚n this country, this wouldn鈥檛 even meet minimum thresholds, it would have just been laughed out,鈥� he says.
Special cameras are used to take fingerprint images that allow some light manipulation to aid legibility, but also keep a detailed data log of all operations so that courts can make an assessment.
For instance, one camera often used by forensics teams in the UK can remove repeating patterns such as those found on textiles from the image of a print. But a log must always be kept to show what changes have been made and to record unedited versions in case they are needed in court.
搁别蹿别谤别苍肠别:听
