Reliable face blurring in videos

Process studies can be a very effective tool to improve efficiency in logistics and production. Taking videos of workers is an essential part thereof – however, protecting the privacy of workers is very important. We at Logivations have recognized this problem and developed a state-of-the-art face blurring solution, which makes individual workers unrecognizable and thus unidentifiable in videos.

Why haven’t others done this yet?

Making people unrecognizable in videos is hard. Once a person’s face is visible - even for less than a second – this person can be easily identified by a human watcher, making the whole effort worthless. Thus, a useful algorithm needs to find faces not 90% of the time in most positions, but all the time in every position. This is where existing solutions fall short. As an example, here are a few results of YouTube’s face blurring technology:

It’s clear to see that the system gets confused by light sources and profile views of faces. Also notable: false positives – regions that are detected as face but really are other things.

Let’s get technical

Here’s how most current solutions work: They use the so-called Viola-Jones algorithm. On a typical face, the eyes are darker than the nose, thus the algorithm looks for such patterns:

Once a part of an image contains such a pattern, it is classified as a face. This works well for faces with no rotation, and as a bonus is quite fast.

But, what happens with faces that can be viewed from the side? Here, only one eye is visible, so the pattern won’t be matched, and no face is detected. Thus, another pattern has to be added for profile faces. This is already quite hard, as there is no simple pattern applicable to every profile faces.

Nearly impossible however is finding patterns for unusual head positions: someone that is looking at the ground, or up a wall. These situations are very hard to recognize with the Viola-Jones algorithm.

That’s the issue: We, as humans, can recognize a known person even if we can only see a profile view of their face. Thus, an algorithm useful for privacy protection has to find all those cases as well and blur them.

Game-changers: Neural Networks

Logivations set out to build a better solution. With latest technologies such as neural networks we used a database of more than 500.000 images of humans to train an algorithm to find patterns for heads on its own – in all positions, rotations and lighting conditions imaginable. Here’s an excerpt of the patterns found:

As one can see, instead of black-and-white pixels, our algorithm found lots of complex patterns. Thus, it can detect heads even in very non-usual positions. This solution easily detects heads where other algorithms fail, such as the one previously mentioned created by YouTube.

In these images, 0.997 means the algorithm has a 99.7% confidence in the detection.

Such an algorithm requires considerable amounts of resources. Using the power of the cloud we can provide a fast and seamless processing of the videos.