Hendon Publishing - Article Archive Details

Tattoo Recognition Becomes Mainstream

Like mugshots, images of scars, marks, and tattoos of offenders have been captured at time of booking for many decades now. These are typically referenced by a text description, so entering the word “scorpion” will filter out just those images that include scorpions. 

While this can be useful, it’s far from foolproof—different officers can call the same content by different names; it may not be obvious what’s actually in a complex tattoo so certain shapes can be easily missed, etc. 

 

While face recognition as a software application is now fairly commonplace as a means of identifying an unnamed face, tattoo recognition is much more complex. This is because a mugshot has predictable characteristics, so the software knows what to look for. 

However, with scars, marks, and particularly tattoos nothing is predictable. The tattoo can be of anything, from a simple small circle to a highly complex full-body work of art, and can be virtually anywhere on the body. This unpredictability has severely inhibited the development of effective tattoo recognition technology until now. 

In addition, different tattoo artists may draw the same object very differently, so when trying to match, for example, a common gang marking, one artist’s view can be sufficiently different from another’s that any two images may well not match.

As computers continue to have ever-increasing power, they’ve reached a level where modern systems now have the number-crunching capability to process the myriad of analyses and calculations necessary to find sufficient unique characteristics to enable one tattoo to be differentiated from another. An example of how well this can now work is shown above.

In the real world, additional issues have to be taken into account. Tattoos, being drawn on flesh, are subject to being distorted, and appear as a different shape when viewed from different angles. In many instances, the tattoo will be partially hidden, which means that the full range of differentiating characteristics may well not be visible. 

So a tattoo can be of anything, be anywhere on the body, be of any size, and may only be partially visible. This is the challenge, and it’s only now being satisfactorily met. Because of the complexity, tattoo enrollment times are longer than those for face or fingerprint enrollment. 

The enrollment process involves analyzing the image to find small areas that have sufficient content, such as lines crossing at certain angles, that are distinctive. The system attempts to find a number of these areas, after which their characteristics are transformed mathematically into a digital string. This string then becomes the unique digital signature of the tattoo.

Once enrolled, database search speeds can be extremely fast. This is because, as with fingerprint or face recognition, the matching process doesn’t involve matching images, but rather compares the digital strings of an unidentified tattoo against the strings of all the tattoos in the database, sorting the results into order of match percentage, and then displaying the images for an officer to visually confirm the correct match. 

It should be emphasized that tattoo recognition, like face recognition, doesn’t display the ‘positive match’ message so loved by Hollywood. It’s more like a Google text search, where the top matches are displayed for the user to make the final call. The actual match may not be number one on the list, but should be near the top.

Because during the booking process a description of an SMT is typically included, a text filter on this field can further narrow down the search, so that even with a partial shot of a tattoo and a very large database, the matching tattoo can be found. Each potential match that’s displayed will include the name associated with it, and this links directly back to the offender’s full record. 

When the text filter “scorpion” is applied, non-scorpions are excluded, as shown here:

 

Since the software sees a collection of distinguishing features rather than an image of a scorpion, it may return images that have similar features but are not of scorpions. A cause of this with tattoos is that an artist may use the same style across a range of different tattoos, with the same characteristics being picked up by the software. Text filters can help address this. 

 

In many instances, only part of the tattoo may be visible. The example here is an extreme case of this, where only the tips of the scorpion’s claws are visible. Again the small partial ‘probe’ image is on the right, and thumbnails of the matching images above the user-defined threshold are displayed on the left. In this example, only one of the images in the database matches the probe above the threshold of 87 percent—that of Duane Berlitz at 90 percent. 

In some cases, the shot of the tattoo may have no boundaries at all, so the system has no information on which to match its position. However, this can still work, as in the example of a tattoo of a tiger here:

 

One of the attractive features of tattoo recognition is that the images of the tattoos already exist in the database, so minimal preparation is required before the system can be used.  All that needs to be done is to point the software at the location of the required fields, i.e., the tattoo image, its description (if this exists), and the name of the offender.  

Once this is done, all the tattoos in the database can be automatically encoded. When this is completed, the tattoos are ready to be searched. So law enforcement gains substantial additional benefit from data already held in the records system, without anything else having to be added.

The tattoos in the existing offender database are accessed in read-only mode, so data integrity is assured. New images added to the database can be automatically detected without any notification from the host application, so the developer of the records system doesn’t have to modify their software to notify the tattoo recognition system every time a new tattoo is added. 

 

The underlying technology can be extended to any type of image, including partial faces and scenes. This is an example of a cropped bedroom shot in JPG format being matched against a full database shot in PNG format taken from a different part of the room. Where hashing and other approaches cannot start to address this type of requirement, image matching technology now can. 

The technology demonstrated here is part of a complete redevelopment of a major system developed for the UK Police, which includes scene, face, and object recognition. This focused on identifying offenders and victims in images of child abuse. It was launched at the House of Lords in London and

earned the

International Law Enforcement Cybercrime Award.

 

Iain Drummond is the CEO of Face Forensics, Inc. and may be reached at iain.drummond@faceforensics.com.



Published in Law and Order, Jun 2016

Rating : Not Yet Rated



Comments 0 Comments

No Comments


Article Images

Click to enlarge images.

 
Events and Tradeshows: LAOPFMTRPSIT
Latest News: LAOPFMTRPSIT
 
Close ...