June 27, 2017 -
In an Amazon guest blog post, Chris Adzima, a senior Information systems analyst for the Washington County Sheriff’s Office, discussed how his team uses AWS for facial recognition as well as Amazon Rekognition as a crime-solving tool to identify persons of interest.
In its initial testing of AWS and Amazon Rekognition, Washington County Sheriff’s Office uploaded 300,000 mugshots from as far back as 2001, which were stored on its jail management system (JMS), into Amazon S3.
Following this upload, the Sheriff’s Office indexed all the images in Amazon Rekognition, which took approximately three days.
The JMS allowed the Sheriff’s Office to tag the shots with a range of information including front view or side view, scars, marks, or tattoos. Since it only needed the front view shots, the law enforcement agency used those tags to get a list of only those images.
In order to upload the images to S3, the team created the bucket and manually used the web interface to upload about 1,000 images at a time.
Adzima admitted that while the process took a while, his team was able to set it up in a relatively short time and allow the automated process to complete over time.
The team then used a PHP script to upload the images from its JMS servers onto the AWS web server.
Once it completed uploading the 300,000 images into Amazon S3, the Sheriff’s Office needed to index all of the images. Adzima said the team later realized that it would have been easier to index the images in the same script that he used to initially upload them to S3.
To index the faces, the team looped through every image in the bucket. The team used the ExternalImageId property so that it knew what Amazon Rekognition would return when it conducted the face search.
Once it indexed all the images, the team created a quick front end (which was made possible by a simple form to a PHP script) that would allow it to search the collection for matches whenever it received a new image.
The team tested the system by running surveillance and other images of known suspects from solved cases and evaluating the accuracy of the results.
In order to ensure that it didn’t taint the results, the team asked a detective to send 20 random pictures of individuals whose identity he knew but the team did not.
After running all the images through the system, the Sheriff’s Office reviewed the results to find the face that they thought best matched the individual.
The team then sent the results to the detective and discovered that 75 percent of the results accurately identified the individual in the image.
After the initial test was completed, the team created a mobile application and web application for officers in the field to capture an image and process it with Amazon Rekognition.
Adzima said Amazon Rekognition successfully identified a suspect based on an image taken with a camera at a self-checkout kiosk in a retail hardware store.
The team used Amazon Rekognition to identify at least two other persons of interest, including a man who used a stolen credit card and an individual who was posting photos on Facebook under a pseudonym, achieving an accuracy of greater than 95 percent and close to 100 percent, respectfully.