For the last six month or so, the Washington County Sheriff’s Department has been leveraging face-searching tech as a way to drum up leads in criminal investigations.
What initially started as a conversation in November of 2016 about how to make the agency’s mugshot database searchable resulted in the December 2016 adoption of Amazon’s Rekognition technology, which layers onto existing databases. Chris Adzima, a senior information systems analyst with the county, said the tech seemed to fit the bill.
“Our team was thinking that there should be a way we could automatically search. We have all of these booking photos and we figured there should be a way we could do that,” he said. “While we were discussing that, about a week into those discussions, Amazon announced Rekognition and we thought that was the perfect thing to test out.”
As it stands, bulletins and persons of interest are sent to Adzima to run through the system for potential matches. The county’s database is home to some 300,000 images, which translates to roughly 200,000 individuals.
Before moving ahead with the project, Adzima said policies needed to be developed to ensure when and how the system would be used in criminal cases and identifying at-risk individuals registered in the Help Me Home database.
“Even though we don’t get any complaints, people can misconstrue what is going on. I think the important thing to know is that we are not out there taking pictures of people and running them to know who they are and where they are,” he said. “These are all situations where a crime has been commited and captured on camera. … Those are the images we are using. We’re not using this for anything else because we don’t have any desire to do anything else with it.”
When it comes to turnaround times and accuracy, the senior analyst joked that the results depend on a couple of variables, including whether or not he is at his desk to run the request. If he is, the requesting party can expect an answer in as little as five minutes.
The accuracy of the system is a little more complicated to nail down. Not only does a decent match require a decent, workable image, but it also requires the subject in question to be in the database. In early testing, Adzima said known offender trial runs registered at around 75 percent accuracy.
For investigators like Detective Tim Kiurski, the tool has been a welcome addition to the suite of investigative resources. In just the six months since the system has been in play, the software has been used in the investigation of everything from a murder case to a simple package theft.
Most often, Kiurski said authorities are working from cellphone photos or surveillance camera stills. “Anybody that can’t figure out who somebody is, we’re going to run it through there as long as it is related to a crime, of course.”
But the software is no replacement for the institutional knowledge of the agency staff or their investigative prowess. Both Adzima and Kiurski agree that it is merely supplemental.
“Sometimes the identification coming from the system is from 10 years ago, so the guy doesn’t even look exactly the same," Kiurski said. "If you were to see him on the street, you might not think he looks the same, but the software is picking up that this is the same guy. ..."
Supplemental though it may be, neighboring agencies have expressed interest and benefited from the facial recognition system, and there is momentum to expand into shared resource with intranet access for county personnel.
“When you are talking about inter-government agencies working together, there are things you have to do and proper steps you have to take, and we are working on that,” Adzima explained. “The idea is that we want to get as many other agencies with these type of images in as possible, because that will expand our ability to help and do these searches. Criminals don’t abide by county lines when they decide they want to break the law."
