The story of how we at Artellence created our face recognition technology for MFIs
In the past few months, the already boring life of MFIs has become even more fun - now, when issuing an online loan, it is necessary to carry out photo verification: to compare the borrower's selfie with his passport. There are quite a few standard models that automatically recognize and check faces in a photo, however, they are trained to work with “standard” photographs. This does not allow the models to take into account the peculiarities of the photos of potential borrowers of Ukrainian MFOs, which directly affects their accuracy.
While creating our own face recognition technology at Artellence, we found quite a few such features. Below I will tell you about the most important of them, as well as share our solutions.
However, most of the pitfalls that affect the accuracy of face recognition algorithms lie in two other planes:
From a technical point of view, the task is much more difficult than the previous one. Although the photos on social networks are of high quality, they were taken from arbitrary angles, often with glasses or grimaces on their faces.
The solution for this problem, which we are now testing, is to compare faces only for features that are significant for our task. For example, give “similarity points” for the same nose shape in two photographs, but not give such points if the mouth is wide open in two photographs of random people.
In the past few months, the already boring life of MFIs has become even more fun - now, when issuing an online loan, it is necessary to carry out photo verification: to compare the borrower's selfie with his passport. There are quite a few standard models that automatically recognize and check faces in a photo, however, they are trained to work with “standard” photographs. This does not allow the models to take into account the peculiarities of the photos of potential borrowers of Ukrainian MFOs, which directly affects their accuracy.
While creating our own face recognition technology at Artellence, we found quite a few such features. Below I will tell you about the most important of them, as well as share our solutions.
1. Photo quality
Fraud, of course, is present in the photographs: borrowers upload scanned copies of other people's passports (more often), take selfies not with their passports (less often), and even send random photos instead of passports.However, most of the pitfalls that affect the accuracy of face recognition algorithms lie in two other planes:
- borrowers take photos of low quality, with glare, highlights, disproportionately small passport photo for selfies;
- also they often show the first spread of the passport, on which they are ... 14-16 years old! And comparing the appearance of a teenager and how it has changed by the age of 40 is still a challenge.
2. Building models
Selecting and comparing faces in high-quality photographs is a fairly simple task and many ready-made solutions have already been developed for it. But there are several important issues that popular models usually don't address:- borrowers can submit photos both horizontally and vertically, which reduces the hit rate of standard models by 40%. To increase coverage, we created our own algorithm for rotating photos at the desired angle;
- glare and light in photos (the problem is more common with laminated id-cards), which “eat up” another 10% of the coverage. It is difficult to deal with this problem at the stage of training the model, since the data has already been lost, much easier - during the creation of a photo.
3. Plans for the development of technology
But we do not stop there, because comparing a selfie and a passport solves only half of the problem and does not in any way identify passports with fake photos. The next task, which we at Artellence are already working on, will be comparing a passport with social networks - after all, a person's life is much more difficult to fake than one photo in a passport.From a technical point of view, the task is much more difficult than the previous one. Although the photos on social networks are of high quality, they were taken from arbitrary angles, often with glasses or grimaces on their faces.
The solution for this problem, which we are now testing, is to compare faces only for features that are significant for our task. For example, give “similarity points” for the same nose shape in two photographs, but not give such points if the mouth is wide open in two photographs of random people.
