Interpretation of image and video evidence

The UK's Forensic Science Regulator has published a short useful document on Forensic image comparison and interpretation evidence. Much of what is says should be no surprise to law enforcement and forensic experts who understand video - but its still a useful document as it sets out some points and useful reminders.

Some Highlights:

If the subject to be identified is a substantial distance from the camera or only occupies a small proportion of the image, then the number of pixels that make up the subject may be low. As a result, the level of detail will be very limited and may not be suitable for reliable interpretation. Enlarging an image beyond its resolution will not add additional detail, but will only duplicate existing pixels resulting in a blocky or blurry appearance, depending upon the method of enlargement. The potential for enhancement of low resolution imagery is generally very limited.

There is a great example of how recorder compression artifacts plus 'enhancement' can lead to bad recovery. Based on the enhanced license plate below you may be quite sure the number plate reads DUN 150J, but the real plate is OUM T5GJ.

 Image enhancement should be approached with caution. If you want to know more about what enhancement or clarification can do, check out Kinesense CEO's post on "Video Clarfication: Fact or Fiction". 

Licence Plate Recovery

The good people at VideoCleaner.com have set a 'license plate challenge', saying:

(Note: Doug asked me to block out the actual number plate in this blog post)

Unprecedented... Hmmm... Spready from Amped solved this challenge easily so I thought, I'm sure Kinesense can do this too using our Video Clarification toolkit.

Sure enough, a couple of clicks and the license plate is recovered:

To recover the plate, I used a histogram edit filter (to crop out some of the added noise), temporal median filter to recombine frames, and rotate to correct the orientation. Then I clicked 'histogram equalization' to make it a bit clearer'. About 5 clicks from start to finish.

However, this is clearly not a 'real' license recovery video. By that I mean, the noise 'mode' is artificial and manually added to create the 'license plate challenge', and would never happen in a real video job faced by law enforcement. That said a more realistic video would be solved using the same clarification filters. 

I dug out this example from my archives. It is a read covert surveillance video, in low light with an older analog bullet-cam connected to a commonly used brand of covert surveillance recorder. The issue was caused by the analog camera's auto-gain going too high in low light, and thus adding a huge amount of noise. Running the 'temporal median filter' across the video recovered license plates of any vehicle parked for more than a few minutes in view, or in this screenshot, it recovers the logo on the skip/dumpster.

VideoCleaner set a neat challenge. Recovery under these conditions is not too difficult with the right tools, and its great to see issues like this discussed and highlighted by the community.