The customer was paying a team to manually review driver’s licenses and photographs against a set of rules in effort to verify age and authenticity of the person in the picture. The goal was to reduce human error and speed up the process by using artificial intelligence to accomplish the same review process.
The Artificial Intelligence Solution:
We began by working with the client to identify in detail the input requirements for the artificial intelligence sequence. We established that the input would include two photos; the first being a person holding up an identification card, and the second would be a picture of just the identification card. As we were asked to create a pluggable architecture that the customer would then add to their internal systems, we identified the output code format in detail.
The overall intelligence process was comprised of two sequences, creating a dual layer of authenticity verification. The first sequence was the facial comparison process, which used artificial intelligence services to evaluate if the photo on an identification card matched the photo by providing a “confidence score” from the four different A.I. services listed below:
- Facial AI Analysis Using Microsoft Cognitive Services
- Facial AI Analysis Using Amazon Rekognition
- Facial AI Analysis Using Watson
- Facial AI Analysis Using Kairos
The process then sorted the scores from these services and created a resultant custom aggregate score displaying the overall intelligent confidence of a match between the person in the photo and the license.
The second intelligence process surrounded verifying authenticity of the identification card used. Google Vision OCR service was called to read the birthday and name of the user from the identification card and was returned to the client code enabling them to compare against the data the registered user had entered. Additionally, the metadata of the photo was analyzed to determine if any editing had been done to the license photo including color channel analysis, EXIF data processing, and image quality parameters. The metadata information was returned specifically as well as returning an overall confidence score that the photo was unaltered.
The architecture of the solution is illustrated below:
Overall, a twenty minute manual process was reduced to a few seconds of processing. By delivering confidence scores from four services, the API essentially delivered four points of view on the confidence in a match instantaneously. That’s like having a four person team work on each case! This is definitely a job better left to the robots.