In addition to sometimes affecting the mood of downloaders, mosaics play a very important protective role more often. For example, for some scenes involving horror, bloody content, or police officers, a mosaic can avoid causing discomfort to people, and can also protect the safety of public officials. Moreover, the coding in the movie is not to intentionally hurt the heart of the otaku, everything is a business need. Generally speaking, there is basically nothing serious about mosaic in daily use.
But the problem is that after the advent of the artificial intelligence era, mosaics may face the risk of being "decoded".
In order to code well, AI is not idle
Although we often come into contact with mosaics in daily pictures and videos, many people may not know its origin.
"Mosaic" is a transliteration of the English word "mosaic". It is a mosaic technique, which is in block shape and originated from ancient Greece. And the way people perform regional blurring of images is to refine the color scale of a specific area, and cause the effect of disrupting color blocks. Such color blocks are like small squares, very similar to mosaics. Therefore, this image blur processing is called mosaic.
In fact, to put it more simply, it is to lower the image resolution. Originally a face can be subdivided into 10 million pixels, but now it becomes 10 for you. Can you still recognize it?
Therefore, when we say whether the code is heavy or not, it depends on how low the resolution is.
Since it is image processing, mosaic naturally basically has two scenes: pictures and videos.
There is basically no difficulty in coding a picture. Since the picture is static, you can select the area and operate it directly. But when it comes to coding dynamic videos, the technical difficulty comes out.
Video is composed of pictures frame by frame. It seems that the most direct way to solve this problem is to encode each frame of pictures, and then combine them together, and the video encoding will be a big success. But it is obviously unlikely. The traditional movie film has 24 frames per second. Even if it is processing a 5-minute video, it has to process thousands of pictures. After the work is done, it is estimated that everything will become a mosaic.
To this end, people have also thought of many ways. YouTube has launched a custom blurring tool through which users can code a series of private content such as their license plate numbers, house numbers, bank cards, etc. On the surface, technical novices can easily put on mosaics. But the embarrassing thing is that this tool can easily expand the code range. For example, if you type a code on the license plate, it is likely that the entire front of the car is gone; if you want to code your face, you may be tired to blur your entire body.
The coding tools developed by Microsoft may be tracked more accurately. The tool is based on artificial intelligence technology: by building a deep learning neural network, it can accurately detect the face in the video to achieve tracking. And in order to avoid "errors" of other characters in the video, it can realize the recognition of different faces. Through this set of procedures, the performance of artificial intelligence coding is not bad.
The use of artificial intelligence coding can be said to be a major advancement in coding technology. People do not need to process image frame by frame, nor do they need to use stupid video processing software to extract, copy, render, and track video clips. To a large extent, it is avoided that the mosaic cannot keep up with the target's moving speed from time to time and lead to exposure.
However, Mosaic came with the purpose of protecting privacy, but it aroused people's curiosity about the real content behind it. AI coding is not idle, and a group of "decoding" people are also busy.
Attacking the Shield of the Child with the Spear of the Child: It seems that AI seems to be better at "decoding"
The logic of decoding is also very simple: Since your mosaic is a fuzzy image processing, can't I just clear it for you?
Not long ago, Nvidia, Aalto University, and MIT jointly developed a technology that uses AI to refine blurred images within a few milliseconds. In fact, there are many similar fuzzy picture processing technologies. For them, it may be useful to process some pictures with lower coding levels. If it is a little more complicated, you will step aside.
Of course, some people are serious about the research and development of mosaic erasing technology.
In 2016, the University of Texas developed a method to eliminate mosaics, which can effectively penetrate the occlusion of mosaics to identify image information. Technicians build a face and text recognition system and use pictures on the Internet to train it. Finally, the accuracy of the face restoration in the blurred video can reach more than 80%, and the face that has undergone severe mosaic processing can also achieve about 50%. The correct reduction rate.
To achieve this level, it should be said that it is enough to arouse the vigilance of those who just type code casually.
Last year, Google came even harder. Just after Microsoft launched mobile coding for face recognition, Google announced that by adopting a new pixel recursive super-resolution technology, Google brain has acquired the pattern of changes between the high and low pixels of the image, so it can realize the coding content. More precise matching. For example, after learning that red is lips, it will restore the corresponding part to lips.
In other words, through learning and training, Google brain can "brain fill" the original face in the photo for the mosaic.
It seems that Google doesn't give Microsoft face much...
I believe that the otaku friends are already eager to try, and I hope Google will quickly push this technology into practical applications, so as not to toss and turn between hope and disappointment again and again.
But objectively speaking, coding is actually an irreversible damage to the picture. Therefore, rather than "restore" the picture, a more accurate statement should be "guess." To use artificial intelligence to restore mosaic pictures in a popular sense, a necessary condition is a large number of picture comparisons. This is true for both the University of Texas and Google, which means that in the absence of original image materials, artificial intelligence can only be infinitely close to the original image.
When the mosaic is penetrated, is our privacy still safe?
However, being close means that it is possible, or not necessarily the same as the original image, but it can be highly similar. From this perspective, the emergence of mosaic removal technology will undoubtedly pose a huge threat to the privacy security issues mentioned at the beginning of our article.
In order to deal with this possibility that no one wants to see, we may consider the following ways.
The first is to encrypt and modify the image. Although the University of Texas has a high recognition rate for blurred videos and coded videos, it only recognizes only 17% of the modified images using the P3 technology (Privacy-Preserving Photo Sharing) developed by the University of Southern California. Correct rate. This technology is developed to ensure the safety of photos spread in the Internet era. There are many similar technologies, which can mislead the direction of artificial intelligence guessing by modifying the details of the photo, thereby protecting privacy.
Second, change the coding format. Researchers at Simon Fraser University have proposed a way to use artificial intelligence to articulate faces that require coding. In this way, it can show the emotional characteristics of the speaker and make the visual effect less dry. After all, always watching a piece of mosaic dangling, the visual experience is really not so good.
Moreover, the face processed by abstract painting should basically have nothing to do with the face... If Google still wants to recognize it, I am afraid that it will take a lot of effort on more complex abstract art paintings. I guess Every ten or eight years may not be able to leave the teacher.
Third, use extreme "coding". The reason why mosaic pictures can be "restored" is that on the one hand, the resolution is not reduced enough, and on the other hand, the pictures after coding will still have a certain amount of original features. So, if you don't want to be recognized, just cover it roughly. Putting a black bar on your face, hitting a white light, or putting up a big emoji if you want to express emotions, artificial intelligence should have no way. But this trick is still used with caution, after all, compared to a big black bar floating around on the video, it is still a bit more comfortable to mosaic.
Of course, we don’t have to be too anxious and nervous in the face of mosaic removal technology. On the one hand, its technical maturity is far from enough, and it is currently limited to the "reduction" of face recognition, and text, trademarks, etc. have not been involved; on the other hand, image blur and clear technology is positive in many scenarios. , Such as the restoration of key pictures in criminal investigation cases, the clarification of archaeological data, and even the renovation of an old photo at home.
Compared with the clear processing of blurred images, it is obvious that cracking the mosaic is the most difficult level. If even the mosaic can be cut under the horse, then the high-definition principle of daily pictures and videos is even more important. Moreover, if the final mosaic removal technology is really mature, since it can easily expose privacy, it will inevitably face legal and moral constraints.
Therefore, under the condition of proper restrictions, mosaics should be played or they have to be played.