I spent a year following facial recognition’s rapid roll-out across the UK for BBC News. So-called “live”or “automated” systems are capable of instantly matching people who walk past a camera against a database of file photos.

For the two police forces testing the technology, South Wales Police and the Met, that database is often drawn from the Police National Database (PND).

The PND isn’t just used for this kind of “real-time” facial recognition. It also has a “hindsight” facial recognition system built in, which police forces can use to look up suspects’ images.

I found a worrying lack of comprehensive data on the PND and its facial recognition system during my research, so I’ve pulled together below everything I found out. Feel free to contact me with additions, corrections and clarifications.

What is the Police National Database?

The PND is a Home Office-run digital storehouse of evidence. In terms of images it includes faces, as well as pictures of things like scars, tattoos and other distinguishing marks. Many, but not all, of these are pictures taken when someone is arrested in connection with a crime. It also includes things like names, locations, companies, but from now on, I’m just going to be writing about the images, not the rest of the data.

Who contributes to the PND?

Police forces submit their images to the PND. According to the Home Office, only nine forces currently don’t upload images to the PND.

How big is the PND?

This is where the numbers get a bit murky. In June 2015 the National Police Chiefs Council said there were 14.2m images in total.

In a 2018 FOI response to me, the Home Office said there were c.21m custody images on the PND.

This implies around 7m images have been added to the PND in three years. Which would mean 6,000 images being uploaded every day, or 4 images a minute. However, this must be seen in the context of duplicate images, discussed below.

How many of the images are faces?

The Home Office told me: “The exact number of facial images cannot be ascertained, as… images, includes marks, scars, tattoos and profile images.”

An exact number may not be possible, but a good estimate of the minimum is. That’s because the PND has a system by which images can be “enrolled” into its facial recognition software (more on that later). The system occasionally fails to recognise a face as a face (for example, if the pupils of the eyes are hidden by hair, or reflections on glasses) but it will recognise the vast majority.

The National Police Chiefs Council said in June 2015 that 13.7m images had been enrolled (ie. recognised as faces by the software). However, In 2018 the Home Office told me it was 12.7m. Either way, the majority of images on the PND are of faces.

Are innocent people’s images held on the PND?

Yes. The images uploaded by police forces also include faces of those who’ve been arrested for crimes (“custody images”) but have subsequently been cleared. Because there is no system for automatically deleting images once the investigation or court case has completed, the PND inevitably includes faces of the innocent.

(By contrast, Police Scotland has a system for automatically deleting custody images. At an Association of Chief Police Officers meeting in April 2016, Sean Byron, a Met Police staff member who was seconded to the Home Office, pointed out that unlike England, Wales and Northern Ireland, the custody systems of Scotland are linked to the court systems, enabling the automated deletion of custody images relating to cases that did not result in a conviction).

How many innocent people’s images are held on the PND?

In June 2015 the National Police Chiefs Council said the answer “cannot currently be ascertained”. (Some of the images are held under a crime reference number, which some people hoped would enable a calculation of the number of innocent peoples’ images held, ie. those without a crime reference attached. However, it’s possible to upload an image to the PND without a crime reference even if it’s related to a crime, meaning a lack of reference number does not equate to innocence).

This issue was addressed in a 2012 court judgement in which retention of images of innocent people in this way was found to be unlawful. The Home Office took five years to do a review, and this is now the rule, according to the Biometrics Commissioner: “Custody images can be taken on arrest and retained regardless of the legal outcome… and can be retained until the police carry out a review after six years or until the subject successfully applies to a Chief Officer for deletion”.

In addition to custody photos, where do police forces get their face images from?

Police forces harvest from a wide array of sources, according to responses to Freedom of Information Act requests that I made.

British Transport Police told me they produce 30,000 still images each year from CCTV cameras.

City of London police said that images “of course can be taken from CCTV and social media as part of a criminal investigation”. Merseyside Police also said they may take images from social media

Durham Police has images captured from local authority and shop CCTV, dashcams, body-worn video cameras and mobile phones (any image not linked to a crime or incident is automatically deleted after 31 days). North Wales Police also confirmed they harvest images from such sources.

In fact, police forces around the country are pulling in images from all over the place. And they struggle to know how many. Images are often stored in crime records, and therefore in order to tell you how many images they have, the forces would have to go through each and every case file.

It’s not clear under what circumstances such non-custody images are added to the PND, and if they are, whether they are covered by the six-year review rule outlined above.

Who has access to the PND?

In 2015 the Association of Chief Police Officers stated “there are 12,000 licences available for PND use”. In July 2018 the Home Office said the PND is available to the 43 forces in England and Wales, plus Police Scotland, British Transport Police, Police Service of Northern Ireland, the Service Police Crime Bureau and the National Crime Agency, a limited number of officers in the Disclosure and Barring Service, Border Force, Immigration Enforcement, Identity & Passport Services, HMRC and the Security Industry Authority. The Medicines and Healthcare Products Regulatory Authority are in the process of signing up to get PND access too.

Is there facial recognition and search on the PND?

Yes. Any of the dozens of forces and agencies listed above can search the national database thanks to the facial recognition function that was added to the PND in 2014.

Just having your face on the PND doesn’t mean you’re on the facial recognition-searchable system. For example, some mugshots aren’t of high enough quality to be searched by facial recognition software. And of course, some PND images are of scars or tattoos, etc.

In order to be part of the facial recognition scheme, an image has to be “enrolled” by the software, which then means it’s been recognised as a face, scanned, and can be searched against.

Tech company CGI, headquartered in Montreal, Canada, maintains the PND and the facial recognition service uses software provided by Cognitec, a video analytics firm headquartered in Germany (not to be confused with Cognitech, a US-based technology firm).

How does facial recognition and search work on the PND?

Users can enter a “probe” facial image, and the software will attempt to match it against “enrolled” face images held in the PND (ie. images that the system has recognised as a face). The system throws up a number of potential matches which, according to NPCC information in 2015, is a maximum of 50. The NPCC said this was not “automated facial recognition”, since there is a still an element of human filtering at the end.

(Most facial recognition and matching systems work, incidentally, by measuring the distance between the eyes. Some also triangulate that with the relative position of the nose).

How well does the matching work?

In 2015 the Home Office’s Centre for Applied Science and Technology carried out the first test on the system, which had already been live for more than a year. Prior to this, according to the Home Office itself: “no other formal accuracy testing had been carried out as part of the implementation”.

The first issue the assessors found was quality – according to their report, not one of the 4,000 sample images met the standards the police themselves had set for decent photos. In many cases the files were tiny (some as small as 7KB), and a third were not well-enough lit to be used. In the end only around 200 images were used in the test.

The Home Office assessors wanted to see how well the PND facial recognition system would perform as an automated system, without users having to trawl the results. They concluded: “out of the initial 211 searches, the automated facial search of PND identified just 20 true matches, whereas visual examination by the tester identified a total of 56 matches”.

How many duplicate images are there on the PND?

Because there is no automatic screening for duplicate images before they’re uploaded to the PND, the Home Office assessors discovered “over 40% of all searches found one (and often many more) duplicate images already in the database”. If applied across the 13m images in the facial recognition system, that would mean 5m images held unnecessarily.

This duplication cost the UK Government money, according to the Home Office report: “The cost of extending the Cognitec licence beyond 10M templates is understood to have been £120k plus an on-going charge of £18k per year thereafter, something that would not have been necessary had de-duplication been built into the enrolment process or, preferably, addressed at a local force level prior to uploading images onto PND”.

A Cognitec spokesperson declined to comment on costs, but said its matching technology had improved since the Home Office report, and that facial recognition results always require review by a human.

How does the PND’s facial recognition and search system cope with different skin colours, or ethnicities?

The Home Office assessors found 18% of those in the 200 sample pictures were non-Caucasian, although it’s unclear from the report how this was decided.

This presented a golden opportunity to test how the facial recognition software dealt with ethnicity. This has been an issue with other facial recognition systems, and again, Freedom of Information data shows that it was something UK police were well aware of.

Minutes an ACPO working group on facial recognition in April 2014 show that Chief Constable Mike Barton noted “that ethnicity can have an impact on search accuracy and that this varies with algorithm”. He asked at the meeting if CGI would consider this and provide feedback to the group, but the minutes appear to show this was never followed up, and the Home Office assessment did not test ethnicity accuracy either.

The Home Office said facial recognition can be an “invaluable tool” in fighting crime. “The technology continues to evolve, and the Home Office continues to keep its effectiveness under constant review,” a spokesman told me. He said work had been done to improve the accuracy of the system since the 2015 assessment.

A spokesman for the National Police Chiefs Council said the technology has the potential to disrupt criminals, but said any roll-out must show its effectiveness within “sufficient safeguards”. Work is being done to improve the accuracy of the system and remove duplicates.

How often is the PND facial recognition and search system being used, and by whom?

Home Office figures show the number of PND facial recognition searches has grown from 3,360 in 2014 to 12,504 in 2017. It was 7,558 from January to July in 2018, so the annual total almost certainly ended up even higher.

Freedom of Information requests show a varying picture of facial recognition search across the UK. Hampshire tops the table, with 5,283 searches in the last four years. In 2016 alone they were putting in 50 searches a week.

West Midlands, Lancashire, Police Scotland and West Yorkshire police forces have also used the search system thousands of times.

And it seems some forces aren’t even aware their officers are utilising facial search on the PND. Eight forces claimed in Freedom of Information responses that they had not used the system, yet Home Office figures show they have made thousands of searches.

Is the PND the only police store of facial images?

No. Forces are able to have their own stock of images. The Metropolitan Police, for example, has a Custody Facial Recognition System that it says has 2.9m images. Unlike the PND, the Met says these are only from custody images.

How is the PND used in “live” or “real-time” facial recognition?

Beyond looking up images in hindsight, at least two UK police forces have tested using “live” facial recognition. This involves harvesting face images from video footage, and checking those images against a database, live and in “real-time”.

A key difference is that, whereas hindsight systems will generally present a user with a list of potential matches ranked in order of similarity, “live” facial recognition generally only presents a user with one; the most likely match. In short, “live” facial recognition needs to be right first time.

In practice, this has meant police setting up cameras (generally either on a van or on lampposts) pulling the live video feed into a nearby control room and using software to match faces on the video against a database of suspects.

Once a match is flagged, officers on the ground will usually go over to the person who’s been identified by the system, and try to verify their ID.

Faces from the live footage are compared to a watch-list of suspects, and that watch-list is often a subset of the PND. But not always. It can include those who are of “possible interest to the police, who presence does not pose any immediate risk or threat to public safety”.

The databases used for “real-time” facial recognition are generally smaller than for hindsight systems – South Wales Police for example used watchlists of between 400 and 1200 people.