The 'Theory' of Mosaic Removal (Mosaic Removal AI): Is AI Not Restoring It? Uncovering the Truth of 'Lost Information' and Unknown Dangers from the Mathematical Limits of Image Processing and Generative AI ◆ 'Mosaic Removal' is Not Restoration. 'Probabilistic Image Estimation from Mosaic Images'

On the 'theory' of mosaic destruction (mosaic removal)

The mosaic used in Japanese adult videos is a process that intentionally pixelates and conceals the genital area to satisfy legal regulations (provisions regarding obscene materials).
'Mosaic destruction' refers to technical attempts to estimate and restorethe information in these concealed areas using surrounding image information, temporal continuity, machine learning, and more.

Basic concepts (theoretical aspects)

A mosaic is an 'intentional loss of information.' Theoretically, the following approaches can be considered:

1. Spatial Interpolation
   A method of interpolating and filling in pixels lost to the mosaic using surrounding known pixels. This ranges from simple linear interpolation to more advanced edge-aware interpolation.

2. Use of temporal information
   In the case of video, since the position and shape of the mosaic may change slightly between frames, this is the idea of attempting restoration by integrating information from multiple frames.

3. Restoration via machine learning/generative models
   An approach that uses models trained on large amounts of image data regarding 'natural human body shapes and textures' (especially recent generative AI and inpainting technology) to fill in missing parts with 'plausible content.' It is important to note that this is not 'accurate restoration' but rather 'statistically plausible generation.'

None of these are capable of 'completely restoring lost information,' and they all have the limitation of estimating and generating from surrounding information. If the mosaic is sufficiently coarse and a large amount of information is lost, accurate restoration is theoretically difficult.

Important notes

• In Japan, distributing or publicly displaying videos that clearly depict genitals may be subject to charges of distributing obscene materials. The act of creating and distributing videos with mosaics removed carries legal risks.

• Explaining the technical 'theory' is different from providing specific methods, tools, or software to actually perform the removal.

+++++++++

'Mosaic removal' is actually not 'restoration'

Mathematically, you can understand it immediately by thinking of it this way.

[Y = A(X)]

• (X): Image before mosaic

• (A): Mosaic processing

• (Y): Image after mosaic

The problem is that different original images (X_1, X_2, X_3...) can create the same mosaic image (Y).

In other words,

[A(X_1)=A(X_2)=A(X_3)=Y] can occur.

Therefore, even if the AI sees (Y),

cannot be mathematically determined.

This is the decisive difference from AI image generation.

The AI

is what it estimates.

Therefore,the appearance of 99% restoration accuracy and the recovery of 99% of the original image's information are completely different.

This is extremely important.

🧠 So, what is the AI doing?

In Bayesian terms,

[p(X|Y)\propto p(Y|X)p(X)].

Simply put,

observed information

+

the knowledge the AI possesses that 'there are many images like this in the world'

↓

the most likely image

It is.

For example, suppose there is a 'human face' behind a mosaic.

AI learns statistical relationships from a massive amount of face images, such as

• eye position

• nose position

• face contour

• skin texture

• hair

• light source

• facial expression

.

That is why

it can generate an image that says.

However, that isnot proof that it was the person's actual face.

💡 A more interesting point

This is where the real problem of 'mosaics in the AI era' begins.

Past:

Present:

is becoming the reality.

In other words,

'Hidden' no longer means 'safe'.

However, this does not mean that 'AI can completely break through mosaics'.

More accurately,

is what it is.

🔥 And the biggest pitfall

As AI-generated images become increasingly realistic,

there is a possibility that humans will no longer be able to distinguish between 'guesswork' and 'fact'.

possibility.

For example,

↓

Human

is what happens.

However, mathematically,

It is nothing more than that.

This is not limited to mosaic removal, but is a problem that exists in image restoration in general, such as

facial recognition, surveillance cameras, restoration of old photographs, security footage, low-resolution video, medical imaging, and satellite imagery

.

Where does the value of 'AI restoration' lie?

Looking at this from a business perspective, it is quite interesting.

1. Technology that truly restores lost information

This is the domain of image processing and computer vision.

2. Technology that generates images that humans perceive as natural

This is the domain of generative AI.

3. Technology that determines 'whether it is real or generated'

What will become important from here on is3.

In other words,

This creates a paradox.

If it were me, I would define it like this:

The term 'mosaic removal AI' is slightly misleading.

More accurately,

.

And, restored image ≠ original image.

If you do not make this distinction, you will mistake the 'plausible lies' created by AI for 'lost facts'.

This is a quite dangerous cognitive bug in the era of generative AI.

And ironically,

Humans once thought, 'If you apply a mosaic, it cannot be seen'.

Now, they are starting to think, 'AI makes it so we can see even what was invisible'

.

But what AI is doing is sometimes not 'seeing', but

'drawing a plausible story where there was nothing visible'

.

The 'Lies' and 'Truth' of Mosaic Removal AI: What lies beyond the hidden information?

Technology that allows AI to improve image quality or make blurred parts look clear is gathering attention. Many people have likely heard rumors that 'the other side of a mosaic can be completely restored'.

However, from a mathematical and technical perspective, there is a major misunderstanding that humans are prone to falling into.

A guide to understanding the essence

1. It is 'imagination (generation)', not 'restoration' The data itself has been erased by the mosaic processing. What AI is doing is not returning it to its original state, but rather the task of creating a new image by predicting 'it must be like this' from the surrounding data.

2. Mathematically, it is impossible to identify a single 'original image' The exact same mosaic image can be generated from multiple different images. Therefore, it is mathematically impossible to prove from a mosaic image that 'this is definitely the original image'.

3. What AI is doing is 'presenting a plausible probability' AI has learned from a massive amount of image data and is drawing based on statistical inferences such as 'there is a high probability that it is this shape under these conditions'.

4. 'Looking realistic' and 'being a fact' are different Even if an image looks precise and authentic, it is merely a 'plausible hypothesis created by AI', and it is possible that it differs from the actual original image.

5. Beware of 'cognitive bugs' in image technology as a whole Similar AI technologies are used in surveillance cameras, restoration of old photographs, and medical imaging. It is dangerous to assume that 'because AI restored it, it must be real,' and it is important to understand the limitations of the technology.

Multiple perspectives and opposing viewpoints

The perspective of those expecting 'restoration'

It is common to think that 'information remains deep within the image data and can be extracted using advanced technology.' When shown an image that has actually been sharpened, people are overwhelmed by its realism and believe they are seeing the 'true form'.

The technical and mathematical perspective that 'restoration is impossible'

Mosaic processing is not a reversible 'encryption' but an 'irreversible process' that strips away information. Since the lost numerical data no longer exists, there is no way to completely recover the original data as a physical fact.

Future impact on society

While this technology is used for positive purposes such as improving the resolution of security cameras and assisting in medical diagnosis, it carries ethical and legal risks where non-existent evidence could be mistaken for 'facts restored by AI'.

The history of mosaic and image processing issues

• 1980s - 1990s: Establishment of mosaic technology To meet privacy protection and legal regulations, mosaic processing, which fills parts of images or videos with large dots (pixels), became common. In this era, mosaics were considered an 'irreversible and safe means of concealment'.

• 2000s - early 2010s: Spatial interpolation and contour estimation With improvements in computer processing power, technologies (algorithms) that smoothly connect colors based on the gradients of surrounding pixels emerged. However, this was merely correcting blurred boundaries and could not bring out hidden details.

• Late 2010s: Use of temporal information in video In video, techniques were tested to integrate information from multiple frames to supplement the contours of a subject by utilizing the phenomenon where the position of the mosaic shifts slightly as the subject or camera moves.

• 2020s - Present: The advent of deep learning and generative AI Technology for 'filling in missing parts (inpainting)' based on massive amounts of data has developed dramatically. This has enabled the output of high-definition images that look as if the mosaic has been perfectly removed at first glance.

• From now to the future: The blurring line between 'fact' and 'generation'With extremely realistic AI-generated images circulating, the problem of humans mistaking 'lies drawn by AI estimation' for 'hidden facts' is surfacing. In the future, technology to identify and verify whether an image is 'created' or a 'genuine record' will become even more essential.

Key points of mosaic removal for beginners

1. Mosaic removal is not 'restoration' but 'AI drawing' It is not restoring lost data, but rather the AI is drawing new pictures to fill in the gaps.

2. When you apply a mosaic, the data is lost Mosaic processing is an operation that strips away information; the original data is not hidden behind the image.

3. Mathematically, it is impossible to guess the original image For a single mosaic image, there are an infinite number of possible original images.

4. AI outputs the 'statistically most likely shape' It draws based on the knowledge that 'if it is a human face, there should be eyes here'.

5. Even if it looks realistic, it is not necessarily 'real' No matter how clear it looks, it is possible that the face or shape is completely different from the actual subject.

6. The Misconception That 'What Was Invisible Has Become Visible' People have a psychological tendency to believe that a clear image is the truth.

7. The Same Problem Exists in Security Cameras and Medical Imaging Caution is also required in fields where mistakes can cause major problems, such as the analysis of car license plates or facial photographs.

8. Mosaics Have Changed from 'Concealment' to 'AI Hints' Mosaics used to be for hiding things, but now they have become 'hint data' for AI to make guesses.

9. From Now On, 'Technology to Detect Whether Something is Real' Will Become Important As AI image generation technology advances, we will need AI that can verify whether an image is a generated product.

10. Be Careful of Human Assumptions (Cognitive Bugs) It is important to discard the illusion that 'AI can decode anything perfectly' and correctly understand the limitations of the technology.