How Does Facial Recognition and Analysis Work?
We get a lot of feedback and questions when customers are looking to choose an identity verification provider. These are things like how does Artificial Intelligence (AI) work with Identity verification? How does facial recognition work and is there a human overseeing the process?
Let’s Start With The Face
Identity verification providers look at a wide array of data to make a claim for identity. But, the most important point is to start with the face. It is unique and genuine to all humans, allowing strong conclusions to be made.
For an image analysis, the computer starts by locating a human face in the field and returns all sorts of “face-related” data. Once a face is detected, the next level of processing is defining the “face landmarks.” Which is exactly what it sounds like. The computer analyzes 27 points like the pupil, tip of the nose and eyebrows (Ex.1).
After the landmarks are found, the really cool work begins. Just a few of the features that can be analyzed are:
- Emotion — A list of emotions with their detection confidence for the given face. Confidence scores are normalized, and the scores across all emotions add up to one.
- Gender – The estimated gender of the given face. Possible values are male, female, and genderless.
- Makeup — Whether the face has makeup. This attribute returns a Boolean value for eyeMakeup and lipMakeup.
- Hair — Though some do not have any, the machine can give levels for baldness and what colors are detected.
- Age — The estimated age of the person in years.
The Rest Is Statistics
Once all the analysis is done and the data is presented, comparing these data-points are a quick job for the automation systems. Comparing each point with a percentage that falls within a studied and programmed, standard deviation helps assert claims.
Just like cutting-edge medicine, this is not an exact science. Facial recognition software is improving and sometimes even needs the human eye to comb over the results.
Human and Machine Join Forces
Artificial intelligence (AI) today is only as good as the humans that are teaching it. Companies today like Google and Amazon have Mechanical Turk programs, which crowdsource humans to process image analysis and tell the machine what that image is.
The thought here is that the more image results a machine has in its’ database, the less chance of something new being seen and a result not coming back instantaneously. Identity documents are the same thing.
A hybrid approach allows human expertise and machine learning to really deliver. Our team of developers are trained to constantly monitor and detect new documents that are our AI may not have seen before. They can quickly classify new documents and images to make our learning models even better. By using the best human experts to train the automation, we are ensuring a constant level of improvement.
Until machines can think on their own, a hybrid with human-machine approach is the best way to continually improve. Experts human that drive industry leading AI, will offer the highest level of surety and results.