Introduction
The Internet is awash with articles detailing the pros and cons of facial recognition technology. These arguments are taken form ethical, political, and cultural standpoints often ignoring the technology itself. While these views are necessary to create regulation that benefits a society the technology is here to stay. With that in mind how the technology can be best implemented becomes a major concern and many arguments relating to the best implementation circle back onto the accuracy of the technology.
AI-Driven
To provide a brief definition of facial recognition technology, it is a sub-set of artificial intelligence (AI) technology whereby a simulated neural network is developed solely to identify individuals from either images or videos. The identification is done by analyzing the facial features of the individual. This is primarily done through a “deep learning” process where the neural network is fed millions of examples to “learn” what to do and importantly what not to do when determining the identity of an individual. In practice, this allows the software to determine certain features like the texture of the skin or even the thermal profile of the face in question.
This is done to either verify an individual or identify them. Verification can be seen when your smartphone detects your face to either unlock or access a banking app. Identification tends to be the use case of the technology that receives the most media attention as it is often the use case most associated with use by banks, governments and law enforcement agencies. Here is where the technology’s ability to identify individuals accurately is given the utmost importance.
Accuracy
In 2014 the error of the technology stood at 4.1%. In April of 2020, the error rate had decreased to 0.08%. This shows a massive improvement in the technology, an improvement that would make it hard to argue against the accuracy of the technology. However, that is not the whole picture and only a snapshot of the entire picture that concerns us at Recfaces. Those above-mentioned numbers were case studies in ideal conditions. What of instances in the “wild” where kidnapping victims need to be successfully identified months or years after being abducted? Here passport photos or mug shots are not going to be compared to similar images but images that will only display a minimum of identifying features if investigators are lucky.
Measures are taken to Improve the Technology
Here is where accuracy tends to jump off a cliff and in certain case studies accuracy fell from 0.01% to 9.3% in terms of error rate. Here shadows, obscuring objects, and subject aging can result in identification errors. In general, this would mean that more stringent measures need to be adopted to prevent errors, and this has been achieved through the adoption of confidence thresholds which measure the prediction against a threshold, say 95%, and if it is determined to be higher than the threshold identification could take place while lower meant the identification failed. It is true that there are issues to overcome but this is being done at a steady rate through acknowledging current limitations and surpassing them. That’s why the key point of facial recognition projects is camera placement and accuracy goes hand in hand with it always.
