Door to CG Vol.31: "Face" that artificial intelligence thinks and "face"

How do humans recognize their faces

The mechanism for people to recognize the faces of others is special, and it is known that brains are different from other objects and shapes.The evidence is different from the brain part used when recognizing the face and the part of the brain that works when the face image is turned upside down (see: Elucidate the brain network necessary for normal face recognition.)

People tend to start to recognize from what they have experienced and familiar in the past.People focus on the elements of "faces" from the various elements they see.The face is also drawn on the play tools for small children, and the face is drawn on the button of the child application.The face is easy for people to recognize and conscious, such as a young child immediately finds Anpanman patterns, advertisements, toys, etc. in the city.

For adults, the cover of the magazines lined up at the bookshelf of convenience stores is all faces of the talent, so have you ever been attracted to your eyes?Conversely, if you have multiple eyes or art works, you may be wondering where to be conscious.

On the other hand, there are about one in 100 types of people who cannot recognize the face of a person called a phase recognition at all.This is not a feeling that you are not good at remembering the face of a common person, and you can not distinguish people without atmosphere, clothes, beards and hairstyles.

Not only artificial intelligence utilization, but also technologies related to the face are used in various situations.There are many other things, including the face of the smartphone camera, the age confirmation of the vending machine, the security of the airport, the entry and exit gate of the office, the personal authentication of the theme park, and the verification of the popular concert to prevent the resale, etc.。

Face recognition and face recognition using artificial intelligence, dedicated algorithms and technology are often mistaken but different concepts.In face recognition, you can find and use your face from images and images.Sometimes you read gender, age, and emotions from your face.Face authentication is used to match the characteristics of the specific individuals registered in advance.It is used as a means of biometric authentication, such as password authentication, fingerprint authentication, and vein authentication with the palm or finger.

When a computer detects a face, it detects a characteristic part of the face, such as eyes (pupil), tip of the nose, mouth angle, cheekbones, and eyebrows.The human recognition recognizes each part of these faces as the whole picture, and unless it is very impressive, I do not remember the details of the eyebrows and the shape of the nose.The evidence is that if you are an acquaintance who knows your face, you will often be able to recognize even if a part of your face is hidden.

It is said that Japanese people read their facial expressions from around their eyes, and Westerners often read their facial expressions from the entire face or mouth, and the reason why mask life in the corona evil was obediently accepted is the expression of the expression from which part.The factors such as are also affecting.

An experience that feels the mystery of the face is that if you make a face image that combines the two faces, you will not know who and who is mixed.Also, even if the wrinkles increase or the color and amount of hair change, the eyes, nose, mouth, etc. are basically the same.It may be derived from such a recognition mechanism to meet friends in the city or to distinguish classmates in the city in the city for several years.

There is a standard face derived from the face you usually see often, and it is said that you remember the difference from there.Therefore, those who have been living in Japan have always been not good at distinguishing Japanese and Asians, but not being able to distinguish Westerners and Africans.

On the other hand, when a computer recognizes or face recognition, unlike a way to distinguish his face, it is similar to the situation when a human painter draws a portrait.In other words, places that are not very affected by facial expressions, such as the interval between the eyes (eyes), the shape and width of the nose, the shape of the chin, and the length from the mouth to the tip of the chin, are used for judgment.As an implementation, the face is converted from 100 elements to a vector expression of about 1,000 elements, and if the distance is close, it is recognized as the same person, and if it is far away, it is recognized as a different person.In addition, even if the lighting and face angle are different to learn a large number of different photos that the same person are in order to increase the accuracy, which part is the same, with the same lighting and face with the same person.Even so, there is also an approach that extracts elements that should be distinguished if it is different, and creates a model with correct face recognition in advance.

When a face is detected from a photo like "Is this you?" In a smartphone tool, such as SNS, Google Photos, and Apple's photo apps such as Facebook, you can laugh at the wrong result.But it will be conveniently used.Currently, Europe is prohibiting remote monitoring using face authentication, and the balance between privacy and convenience will be discussed in the future.

For example, the use of a lie founder using AI has become a hot topic, but in order to indicate that it is not recognized as biased, the tool kit that reads emotions from the face image of Emopy is available in all sourcodes.doing.Just because the source code is published is not a proof that there is no bias.However, compared to those who do not know the contents in a black box, it is a sense of security that everyone can check.

CLEARVIEW AI, which collected information for face recognition from photos publicly published on SNS, is controversial, so can I do anything if I just follow the law?That is not the case.Transparent data is transparent so that the face photo data is used unexpectedly in a place where the person does not know, it is reused for applications that are not permitted, or the data is not transferred without permission.We believe that it is necessary to keep gender and the basic rights of people and keep considering the unreasonable treatment.

Utilization of non -existent faces

The face made of CG is approaching a real human by expressing wrinkles, spots, pores and scars.Conversely, real human beings are desperate to remove face wrinkles, spots, pores and scars.

If it is a real person's face, it is difficult to collect large amounts of data in an ethically unprecedented way, so it is used for artificial intelligence research, face recognition, and face recognition tests.There is also a flow that uses non -existent face images.DataGen and Synthesis AI can generate a large amount of face image for research, manage the faces of non -existent people with parameters, and collect facial materials from the desired attribute.

"Door to CG VOL.30: SIGGRAPH2021 Report "Fight with Deep Fake" "In this Person DOES NOT EXIST, you can see unrealistic facial images one after another.Occasionally, a strange face image comes out, but in most cases it can't be distinguished.Humans can feel the face with a little discomfort, and it is said that the face image made by artificial intelligence can be distinguished because the eyes are distorted, but the performance is enough to surpass the distinction level.What is coming is the current state of machine learning.

On the side of the WHICH FACE IS REAL?, One is a photo of a real person, and the other is a photo of an unrealist who created by artificial intelligence, which face is real?I will give you a quiz.Most people are indistinguishable, and whether they are right or off, they have no choice but to rely on the "feeling" that is not the answer.

This "no existence ..." has various different versions, which can be said to be a parody version.

Words that do not exist: https: // www.THISWORDDOESNOTEXIST.COM/

Cats that do not exist: https: // Thiscatdoesnotexist.com

Dog version that does not exist: https: // gitHub.COM/GuillaumeMougeot/DogFaceNet

Non -foal character version: https: // thisponydoesnotexist.net/

Non -existent animal character version: https: // Thisfursonadoesnotexist.COM/

Non -existent anime version: https: // Thisanimedoesnotexist.AI/Non -existent anime character version: https: // www.THISWAIFUDOESNOTEXIST.Net/( * Derived from "bride" wifes → waifu who likes anime lovers around the world)

GAN (hostile generated network) and Stylegan are used for these productions, but it is interesting that cats, dogs and anime characters have a slightly different approach.The cat version is based on a part of Cat Dataset of more than 9,000 cat images in Kaggle.In the case of dogs, the balance of the eyes and mouth vary greatly depending on the varieties, and many dog photos usually open and give a velo, so it seems that there were differences and difficulties from humans and cats.am.

As a slightly unusual approach, there is a paid service called Artbreweder, which generates a "unrealistic face" from a single face photo.Various processing is possible, including portrait style, CD and record album jacket style, anime -style characters, and characters that are likely to appear in the latest 3D games.As the name of the service, you can make your appearance gorgeous, change the atmosphere, and focus on various factors as you feel like a pet breeder.

The role and importance of future face recognition

Business policies related to face recognition and face authentication vary from company to company.Among technology companies, Amazon and Microsoft have stated that they will not provide facial recognition data easily, and IBM has stated that it will withdraw from the face recognition business.Apple also handles the framework of face recognition.Facebook, which uses face recognition for SNS applications, has revealed its use policy and states as follows.

1. 他人になりすましたり、他人の情報を不正に利用することはできないこと
2. 希望により顔認識機能をオフにすることもオンにすることもできること
3. 顔認識によって他人に身元情報が漏れることはないこと
4. 18歳以上のみが顔認識の機能が使えること
5. 顔認識によって個別操作によってタグ付け機能は行われる（手間はかかるが代替手段が提供されていること）

Although it is easy to focus only on various negative aspects, the Facebook's face recognition function can benefit from the technology that the visual impairment can also benefit from alternative textbooks on whether or not they are in the picture.There is also.On the other hand, Russian SNS VK.COM has set up a function to find a SNS account of a face in the photo from photos taken in the city.These tasks, which have different ideas and different ways of thinking, are not right right now.With the evolution of technology, it is necessary to be aware of it as an issue.

The face is the same as the fingerprint, and the shape and characteristics cannot be changed unless it is good.Therefore, it is necessary to have sufficient consideration and sufficient mechanism to use it in authentication.In the Internet world, IDs and passwords have been leaked, so it may be re -attached to a new password just in case, but since the face authentication code was stolen, I changed my face and fingerprint just in case.I can't.

As you can see from the technologies of face recognition and face recognition, the use of artificial intelligence is suitable for human brain and approach, overwhelming computing power and machine learning, and the human brain.Currently, different approaches are mixed with different approaches and results.In other words, I think that hints may still be hidden in the brain mechanism and structure.

Google Jeffrey Hinton, who won the Turing Awards in 2019, overcomes the task of neural networks and predicts the next technology."How to Represent Part-Whole Hierarchies in a Neury Network" (Abstracted: How to express a part-time hole hierarchy in the neural network) proposes a new model called GLOM ( * Glom is an acronym.It is not the abbreviation of AGGLOMERATE [lump] and the colloquial Glom Together [Kick and line up]).

GLOM has not yet implemented an implementation and is only a discussion in the dissertation, but various approaches with artificial intelligence continue to evolve every day, and the current common sense has become a past, and new ways of thinking and new usage are.It may be an interesting place to spread and it may be a difficult place.

Future schedule of this series: "The door to CG" is a little different from the mere AI topic, and we will cut out topics from the context of CG/VFX and art.Please look forward to the use of AI in the field of video production, advanced tools that have increased value with AI, topics that make you feel the potential in the future, and technology.If you have any theme or wishes you want to take up, please let us know the editorial department.

Contributor: Koo Ando