Is there anything more annoying than a pixelated photo? The truth is that there is nothing worse than wanting to see an image, especially a face, enlarge it and realize that the only thing you have in front of your nose is a set of fat and annoying pixels , ready to completely hide what you wanted see. Well, Google wants to find the solution to this problem and has achieved it thanks to the Google Brain project . The team has just presented a document called Pixel Recursive Super Resolution that would be the name of this new technology and that would summarize an investigation carried out with the purpose of discovering what is behind a pixelated photo. The purpose of the tool would be to improve andadd quality to images that are pixelated or have a very low resolution. But how does this tool actually work? In which fields could it be most useful?
The images you see above are the ones that Google used to perform these tests . The photograph that is pixelated is 8 × 8 pixels and as you can see, the faces that appear are impossible to see. What Google Brain does is use artificial intelligence (AI) to fill in pixels and improve quality. And although the image resulting from this process is still clearly distorted , very important details can be seen, such as that it is a person, the color of their skin or even the type of hair.
It is actually about making an approximation of what the original image would be . The result is not completely real, but rather a kind of recreation generated from the first pixels. What this software does (yes, it is nothing more than a computer program) is to use two different neural networks that generate images. The first is responsible for mapping the pixels to equate to a high resolution image: in this way, you can set a "skeleton" of the real image, with the structure of a face or the composition of a room. The second network is the one that takes care of the rest. This adds new details, which are based on other images that the system has stored and that havepretty similar pixel settings . Thanks to these two technologies, Google can create a final image that recreates part of the details that could be real. However, the Google Brain team will still have to work a lot in this field, so the results that we will obtain in the future can be truly impressive and, more importantly: useful.
Although it is clear that this is not a system that can be useful to identify suspected criminals (the mistakes that could occur would be flagrant, at least in this very early phase of the project), it could be useful to validate certain details or clues, such as some important feature of someone's physique or defining image objects that simply disappear when we enlarge a photograph too many times. It is clear that the results will not be as accurate as we would like, but they will look much more dignified and presentable.