Hello, my friends~ I am Shi Zhong. My daily life is a big technology giant who is flirting with all over the world. I will try various postures and tell you their endless imagination and warm stories. If you want to be friends with me, you might as well add WeChat (shizhongmax).
Quark "Ferrocious": a wild adventure behind a mobile phone lens
text |
in history 1 Do you know what the village where Li Xiaoyao was born?
I know that it is called Sheng Fishing Village, located in Yuhang County, Hangzhou. It is remote but has beautiful scenery. . .
Why do I remember so clearly?
Because when I was 10 years old, when I first played "Legend of Sword and Fairy", the plot had just begun. Looking at the enthusiastic villagers and listening to the warm BGM, I suddenly had a strange idea: Why can't I, Li Xiaoyao, be cowardly in my hometown for the rest of my life to pick wild fruits by playing wooden swords in my hometown, and I have to go outside and be beaten by monsters?
But then I thought about it, I would not encounter Zhao Linger if I did not encounter monsters.
Thinking of this, DNA moved. I knew that there was a tiger in the mountain that was biased towards Busan lines, so I walked out of the novice village.
OK, we may not have played the same game, but we all live the same fucking life. You and I have (have to) set out from the novice village, when facing the test papers like heavenly books, and when facing the boss's deadline of "tonight", they will fantasize: How good would it be if a divine sword could fall from the sky, and a level 999 of each sword would be. . .
Today, Brother Zhong will tell you a story about the "Divine Sword".
(I) "War Wuslag" and Scanning King
"My child can't find the paper I just sent before the holiday! I had to beg my classmate's mother to take a photo of the test paper for me in the parent group and print it out for the child to do it. But the child has already written the answer, so I can only use a note to block the answer and copy it... "
This is a post posted on the Mother Gang forum two years ago.
such an incomprehensible plot has attracted inappropriate resonance, and hundreds of parents replied: Ah! I thought only my child was like this. . .
It seems that there are more than one trash lurking around you.
There is a person lurking among the melon-eating crowd. He wears glasses and looks like a student. It is not logical that there is a "hard-hard" website like Mom Gang in this world.
This person is Yin Yibin.
Yin Yibin
Yin Yibin is the product manager of Quark App. The reason why he wanted to visit the parent forum was because he was responsible for the "learning section" that was popular at that time.
He is also a warm man. "How can mothers suffer like this?" So I quickly told my friends about this discovery.
A month later, Quark App added a small function - handwriting test paper - to take a paper, which can automatically remove the handwritten words inside; not only that, it can also automatically turn the angle of the paper, turn the paper white, make the words clearer, and print it out more than the original version. The academic master next door was so greedy. . .
is this effect⇩⇩⇩
Although it is such an inconspicuous function, its popularity is hard to describe, and the number of background calls is like a rocket that goes straight into the sky.
In case some friends are not familiar with quarks, here the Zhongge inserts them strongly and introduces a few more background postures:
Many people think quarks are a search app, but in the minds of young users, search is just a dish in the "functional meal" of quarks. They can also use quarks to read novels, watch movies, take pictures and recognize everything, document translation, resume generation, selfie ID photos, PDF to Word, screenshots, transfer to gifs, register to buy medicine, learn recipes and check express delivery. . .
I think the dream of quark is to become the " Swiss saber" in everyone's mobile phones - no car, no house, no women's votes, such a big thing that cannot be solved, but when encountering a small and clumsy difficulty, you can take it out and get it done.
Show you the list of quark tools⇩⇩
In order to achieve this goal, not only Yin Yibin, but everyone in the Quark team has a "needs collection habit". . .
They are like octopus: the product manager does research, the technical master reads the data, the operation lady receives customer complaints, and there are various channels to communicate with users anyway.What users want can be passed to the Quark team in minutes. Maybe the new feature will be made in a few weeks, and it feels like being "poisoned".
To be honest, in today's world, there are not many Internet products that can mix with users such good brothers (good sisters).
, but this is easy to understand, because the Quark team is as young as their users, almost a post-90s generation. The youngest one was born in 1999 when Macau returned to the country. After all, Chinese people don’t lie to Chinese people, and young people don’t cheat young people.
OK. After meeting these young people with thick hair, we return to the main plot.
created "handwritten test paper", and countless children can curse and do their homework happily, but this is just the "Prologue" of the story.
quark This group of young people raised their noses like Hu Bayi , and suddenly felt like they were standing in front of a treasure gate that was "not yet fully developed". This treasure is called - Scan. The reason why
has not been fully developed is that scanning this treasure has been discovered long ago, and there are some scanning products on the market.
is just a group of quarks. This group of people think that their predecessors did not dig carefully, or that their exploration technology was not in place, so they took the specific task of "handwritten by the test paper" as an example. No one did it so meticulously before them. . .
So, an internal discussion unfolds: If you are the best scanning tool on the market to scan "paper files" into "digital documents", who agrees? Who objected?
The classmate from the 1999 team quickly raised his hand:
When I was in college, oh, that is, last year, I really wanted a tool like this. Because the teacher's PPT was flipped too quickly, I could only use the camera to take pictures one by one, and then organize it using the memo function of iPhone . Although rounding is useful, people like me who are socially terrified, sit at the edge of the classroom, and the photos they take are trapezoids. . .
Before he finished speaking, another classmate interrupted:
My cousin works in , an investment company in . I remember him complaining that he had to scan and archive various paper documents and documents every day, and the scanner is difficult to use.
at this time, the horse was reported. In the user survey of Beijing, Guangzhou, , Qinhuangdao and other cities, many of the main users of quarks - teachers and students - also complained: it is best to learn knowledge by learning it, you have to copy the wrong questions and practice it repeatedly. But unless you love calligraphy, who can be patient to copy so many wrong questions by hand?
Moreover, humans have entered the cyberpunk era. It is too old school to spend so much time on physics copying questions? How great is it to have a convenient and powerful mobile phone scanning tool!
As the saying goes, if you want to do a good job, you must first sharpen your tools. Do you want Harry Potter to cut Voldemort ? Don’t you give him a broom first? If you let Xiaozhi kill the Pokémon all over the world, don’t you give him a Pokémon first? If you ask your child to take the Tsinghua exam, shouldn’t he be given a “scanning king” first?
Quark Scan King Product Director Xia Yatong slapped the table: I decided! Let’s help everyone make the “Scan King” better!
Xiaozhi, ask yourself, if the Pokémon Ball is changed to a three-bedroom and one-living presidential suite, would Pikachu be unwilling to come in?
(II) High-definition uncensored large image, get ✓
When it comes to scanning, the scans we saw when we were young were like this: press paper files on a scanner, and then "scream" and the exact same picture is displayed on the computer. The operation of
is called "optical scanning". In fact, it is already an old almanac.
Listen to you to popularize science. The current scan is even better than girls’ makeup:
Step 1, Take pictures
Use your mobile phone to take a photo of the file, and the app will automatically help you cut, set the angle, and remove useless information. It is equivalent to applying a concealer foundation and uploading the picture to the cloud.
Step 2: Revise pictures
Cloud computing power is abundant. AI can identify defects on the file (such as dirty spots, watermarks, handwritten traces on the paper) and remove them; then identify useful texts, deepen them and become clearer. It's equivalent to putting on red lips with full makeup.
Step 3: Transfer this picture to the App on your phone in the cloud. The App will help you export the makeup photos, that is, the picture of the scan result.
ends perfectly, spreading flowers out of the street. After three steps in this way, it is so high that it goes nowhere than the "plain face facing the sky" of optical scanning.
However, if you want to wear a crown, you must bear the weight. This "beauty journey" is much harder than most people think. What is the difficult method of
? Let’s first invite the first player from the Quark team to speak out, he is Wang Huai. Wang Huai’s R&D team in Quark App is responsible for the “first step” we just mentioned.
Wang Huai
Wang Huai is an "obsessive-compulsive disorder". You can see that even if you hang a painting with a skewed 1 degree. But it takes this kind of meticulous spirit to be an App client.
just received the task, and a question that was not a problem was blocked - what method should I take pictures?
You must have never noticed that there are actually two ways to obtain pictures on the camera:
One is fast but the image quality is touching - it is to enter the camera mode and take a frame from the flowing picture that can be seen just by just making do.
One is that the image quality is good but the speed is touching - it is to enter the photo mode, focus honestly, adjust the aperture shutter, and take a photo.
Wang Huai is the first type of "video screenshot". Because they used the "Recognition of Everything" function before, they were using this method. Why ? Because those pictures are used for "categorization" - for example, if you take a dog, even a vague dog, AI will think it is more like a dog than a cat - so it doesn't need to be clear, taking it fast is the best way.
, but scanning is different. Its ultimate goal is not to classify, but to clearly present every detail in this photo.
There is no way, Wang Huai can only choose the "photo mode" honestly. But in this way, the experience is not very smooth, and he made 10,086 optimizations with a group of obsessive-compulsive disorders.
You can take photos next! Wang Huai turned over his phone and looked at it. Damn it? ! Why are there so many cameras?
The second question came to me again - which camera should I use? What are the words of
? Wouldn't it be enough to choose any one?
"No. We tried it. telephoto lens is generally large in aperture, and the depth of field of the photos taken is small. In this way, if the user's shooting angle is tilted, it will cause half of the clearness and half of the imaginary focus... wide-angle lens is generally small in aperture and can be photographed clearly. But if the lens field of view is too wide, the paper accounts for a relatively small proportion of the picture, and there are not enough effective pixels... "
Wang Huai explained.
is like this⇩⇩⇩
They finally wrote a set of rules, asking the Quark App to "Tang Bohu to Light the Autumn Fragrance" on their mobile phones to find the most suitable camera for scanning.
can choose Qiuxiang in actual combat? Wang Huai was a little murmured. . . He also asked the students from the test team to take thousands of photos with different models of mobile phones, and then adjusted each model before he could finish it.
can take pictures now, right? Not OK. . . There are hundreds of millions of users of
quark. You can never imagine whether you are catching the moon in the nine heavens or catching turtles in the five oceans when taking photos.
For example, some people tremble their hands, and the photos they take are Parkinson's imitation shows; some people are absent-minded, and the photos they take are slanted to doubt their lives; some people are dark night knights, and they don't turn on the lights after taking pictures of things.
So the third question is - how to ensure clear photos?
Wang Huai and his team added a reminder on the photo interface - when they detect insufficient light on the picture, they prompted to turn on the light; they also added a grid line - if the eyes are broken, you can refer to the grid to take horizontally and vertically; they also made a real-time focus module - even if the user suddenly presses to take a photo, it can ensure that it is as clear as possible.
Seeing this, a brother may ask: What kind of trivial matter are these?
Actually, I deliberately chose these trivial details to tell you. Because you know, it is difficult for people who can be so "obsessive-compulsive" in the face of these details to cut corners on more serious matters such as image cropping, image compression, and image upload links.
"Swish——", as the picture was posted to the cloud, the heavy baton was handed over to the second player, and the cloud algorithm engineer Huang Ruihua.
Huang Ruihua
Huang Ruihua is an experienced driver of technical control. His driving style is: "I can't have a car in front of me."
If he encounters any technical difficulties, he has to find all the solutions from all over the world if he doesn't sleep, and then he will not allow any solution to be better than himself. . . During the development of
, Huang Ruihua frowned and looked at the pictures uploaded on the mobile phone: Hello, why are there many of these pictures "S body" ? ?
Think about it, if a piece of paper can be spread flat and let you take photos, it is a situation where you burn high incense. In most cases, when we open a book and take a page, the page will definitely be bent, and the closer we are to , the greater the degree of bend, just like the black hole of Hawking . . .
Huang Ruihua was furious at that time. A piece of paper is so "sexy", which greatly affects the scanning effect? !
The next question is - how to straighten the S-shaped pages?
In fact, the method is to use artificial intelligence to let the AI face the bent page "YY" what it should look like when it is flat.
But the problem is that if you want artificial intelligence to learn this skill, you must first give it the learning materials - "comparative data" when the same piece of paper is bent and flat. This data cannot be just one or two groups, it needs thousands of groups.
Huang Ruihua explained to me.
Although, where can I find so many book pages of crooked data? This data is not Coke, you can buy it if you find a store on the street. . . And it’s too late to take pictures one by one by one. When you finish shooting, college students waiting to scan the pages should receive their pensions. . .
A classmate who usually likes to play games thoughtfully: Hey, don’t you use physics engine for 3A game masterpieces?
reminds the dreamer in one sentence. Huang Ruihua organized everyone to create 3D models of several books, and then used the game physics engine to "ravage" it in various ways. Isn't this a lot of training data all at once? This is the effect of
.
As the AI book page straightening effect becomes better and better, Huang Ruihua whistled to the sky: Sure enough, an AI engineer who knows how to use 3D modeling is not a good driver!
It is already the spring of 2021 after finishing these.
The ugly daughter-in-law always has to meet her parents-in-law, and the "Quark Scanning King" has finally been released.
Spring River Water Warm Duck Prophet, a Hardcore Fan Prophet, a New Function of Quark, a Hardcore Fan Prophet — said that time, in Quark’s fan group, sharp-eyed people have discovered the Scanning King.
Quark team moved small stools and sat in a row, waiting for the praise to come, but the praise did not wait, but the board bricks came.
"Why is it a flower when I scan it out?" a fan asked.
When Huang Ruihua saw the "evidence" sent by his fans, he actually took a photo at the computer screen, which was full of molar pattern . . .
"Why can I still see the words on the back after scanning this paper?"
Huang Ruihua asked and realized that this guy's paper was too thin, it was simply a translucent lace style, and the words on the back were penetrated without any pressure. . .
"Why do I still have my fingers when I scan it?"
Huang Ruihua looked at it and found that this guy was scanning with a book with his hand, and his fingers were also taken in. The AI thought it was an illustration in the book, so he kept it for the strong. . .
"This is easy to do. Just tell them not to hold it with their hands when taking pictures?" I suggested.
"The user must have to support him with his hands. Since we want to make the best scanning tool in the world, it is impossible for users to adapt to us." Huang Ruihua looked generous.
Based on the principle of "being the most awesome scanning tool", the masters continued to rework: they optimized the artificial intelligence algorithm to better recognize the words on the back; they also taught AI to recognize fingers that were randomly penetrated at the edge and erased automatically; they also studied the filter for removing molar patterns.
Note that the various filters mentioned just now are "buffed" layer by layer.
For example, the output result of the molar removal of the previous layer is the input information of the next layer to the edge of the finger; the output result of the output result of the next layer to the back of the finger is the input information of the next layer to the back.
It's very similar to an acrobat stepping on stilts, with a water tank on his head, throwing 12 eggs in his hands, and 5 hula hoops on his waist - if there is any problem in any link in the middle, the entire acrobatic will collapse. It is difficult to scan
, and it is difficult here.
The straightforward Quark user is a clear mirror. If you are not satisfied, you will complain, and if you are satisfied, you will praise you. After several improvements, the Quark fan group turned against the wind and became a praising group.
Yin Yibin still remembers that a user came across and said, "I have never used such a good scanning product. It's so white, just like a newly bought book!"
But at that time, the team of scanners was complacent, just like Elder Tang who just released Changan . I didn't know that there was a group of monsters waiting for them around a napkin.
On that day, the operation lady found Yin Yibin and mentioned a user complaint to him. The user of
left a message saying, "I can't use the file scanned with quarks!"
"It's so white, why can't it be used?" Yin Yibin muttered, and decided to call the user directly.
He suddenly realized that the user was scanning an appraisal report and under the report, he specially laid the anti-counterfeiting watermark of the issuing agency. But in the eyes of AI, aren’t dirty things like watermarks removed? It wiped clean in one second and said to the user: praise me quickly!
user said: I thank you!
is unique, and another user's feedback also comes in: he wants to scan an invoice for reimbursement. However, the invoice scanned by the Quark Scan King is too white and tender, and after the filter of "clearity supplement", the red seal will become dim, and it looks like a fake ticket bought on the street corner. . . The guy didn't dare to submit reimbursement. . .
It turns out that "one white covers a hundred ugliness" is not always the truth.
A huge philosophical question is placed in front of the Quark Scan King team: When "truth" and "beauty" sit at both ends of the seesaw, how should we choose?
is the same as life. No one always needs reality, and no one always needs beauty. So our conclusion is that you need to understand the user's true intentions at that moment of scanning in order to serve him more accurately.
Yin Yibin explained.
So, in the subsequent upgrade, Quark Scan King added an "automatic recognition logic": judging that this is a document, it mainly focuses on "real presentation"; judging that this is a document, it mainly focuses on "black and white clear".
This thinking has far-reaching influence.
Since this incident, the Quark team is like a treble speaker in his ears, shouting: "Are you sure you really understand what the user wants to do? R U F**KING SURE?!"
"The alarm bell keeps ringing", they have a surprising discovery--when the user scans, they sometimes don't want to "scan".
(III) When you scan, what are you "scanning"?
"In our lives, we are not afraid of meeting love and sex. What is rare is meeting understanding." This is the line written by Liao Yimei in the drama "Soft". The sentence
is a perfect metaphor for the difficulties of Internet products.
In the real world, even if you have a long conversation with your best friend, you cannot fully understand his intentions; not to mention that in the Cyber world, the system providing services is thousands of miles apart from the user, and can only "guess" through the puzzling fragments of the users.
I ask you, when a user scans a string of express order numbers on the computer screen, what does he want to do?
Let me ask you again, when a user does not call the camera, but goes to the album to retrieve a picture with famous quotes and aphorisms and enters the Scanning King, what do he want to do? Is there a possibility of
? At this time, they do not want to "scan", but want to extract the text in the photo and copy it to another place for use?
may or may not be. . .
In the summer of 2021, Quark, a group of young people, launched the "text extraction" function with this suspicion.
This operation effect explodes.
Judging from the surge in background calls, the user seems to be saying: Such awesome function, is it just now?
Why do people have such a special liking for text extraction? I really thought about it.
Just look at our own brains and you will know that the human right brain is responsible for image thinking and the left brain is responsible for logical thinking. Long-term evolution has caused the brain to have such a division of labor, which must be because it is better than "a clump of left and right brain". What are the benefits of
? Save energy.
Language and text belongs to logical thinking - "abstract information" can be used to reflect "concrete world" - and abstract information is really place-saving.
You see, it only takes half a second to say the word " apple ", and it only takes two seconds to write these two words, but it may take one minute to draw an apple. When you meet a clumsy person, you can't help but consume two apples' calories when you draw an apple. . .
Therefore, the essence of the "text extraction" function is to extract the "logical information" (rather than visual information) in a "image", copy, paste, modify, and search as you urinate and defecate - greatly reducing the cost of information processing.
directly added a logic tool next to the original visual tool, submachine gun + mortar , buy one get one free, it will definitely be popular!
(I have experienced text extraction for you. I am forgiven for being in a hotel at this time, and I don’t have much paper to find...)
After tasting the sweetness of the “extraction” function, Quark and the “gold-touching captains” seemed to have found a new secret door in the scanning field, advancing to the deep secret realm. . .
Yin Yibin recalled that when the product team sorted out user feedback that day, they discovered another strange case.
This user feedbacked a handwritten shopping list, which read: Spinach XX yuan, bought XX jin; cabbage XX yuan, bought XX jin, and so on. The user's request is: I took this shopping list, can you convert it to an Excel form for me?
It turns out that he is a hotel buyer. He purchases goods every day and takes stock at the end of the month. He has to add the 30-day data amount. In the past, I had to add up the handwritten numbers for half an hour, but it might not be correct. If there are tools to turn the account into Excel, then you will definitely go directly from Stone Age to Industrial Revolution !
Yin Yibin saw it and it made sense! He went to find a technical master and asked guiltily: Can handwriting be extracted?
Huang Ruihua laughed: You can try it.
In fact, this technology Huang Ruihua team is not that strange. In the Quark's "Learning" module two years ago, a small function was implemented: converting the student's handwritten answers into printed matter.
But note that the difficulty of the two is different.
A student’s purpose is to let the teacher who judges the paper understand it. There is a lower limit on the degree of sloppyness.
For example:
A person writes a purchase order just to keep accounts for himself. There is no lower limit on the sloppy level!
For example:
In short, the handwriting recognition technology I did for the search questions before was not good at all. We must increase the data dosage and retrain.
However, this time Huang Ruihua met a strong opponent. The deeper the training, the more crazy the technical team will be - as if they are in a bottomless pit. . .
The difficulty of handwritten is not only sloppy, but also lies in the complex meaning of handwritten information.
For example, when registering a visitor, the person above fills in the gender column "Male". When you fill in the form, type two points below to indicate "Same as above". Artificial intelligence can't figure out what humans can easily understand.
For example, if we have written a paragraph, we suddenly want to insert a new sentence between two sentences, and we will type a caret and then write the new words in the gap. This is a meaning of AI, which is difficult to understand.
Huang Ruihua explained. The hell level situation in
is still behind - handwritten and printed words appear on a piece of paper at the same time. . .
Give some tips:
For example, I wrote "one" on a fill-in-the-blank question, but in the eyes of artificial intelligence, this "one" is a horizontal, and the horizontal bar below is also a horizontal. Isn't this "two"?
For example, I wrote "women" in the grid book, and artificial intelligence will definitely question you, what is this TM not "Nanqiu"?
Use heavy dictionaries in troubled times, and difficult problems must be ruthless. To solve this problem, the master sent three AIs to fight:
First, train an AI, which is not responsible for identifying text, and is only used to distinguish which stroke is handwritten and which stroke is printed.
, then, is normal AI text recognition. Since the two strokes have been separated as much as possible just now, the recognition accuracy will be greatly improved at this time.
Finally, I sent another AI to read it in combination with the context. For example, "a beautiful day" is obviously more smooth than a beautiful two days; "she is a woman", it is more right than "she is a prisoner". . .
In short, two months later, when Huang Ruihua showed the handwritten text extraction demo at the meeting, Yin Yibin jumped in fear: Brother, you really made it!
With the two text extractions of "printed style" and "handwritten style", the King of Quark Scanning is somewhat worthy of the title of "king".
Think about it. Whether you see sentences in a book, slogans on the street, or text embedded in the illustrations, you can calmly take out your phone under the gaze of your friends, copy the text in one second, and send it to him "swish--" and your hands will not be dirty.
This feels cool. Installing X-powered tools is not necessarily a good productivity tool, but a good productivity tool must be a powerful tool for installing X.
handwritten text recognition accuracy rate reached 99%, and handwritten text recognition rate reached 90%. The Scanning King team thought that things had been done. But they were wrong again. . . Users are like a box of chocolate candy , you will never know that the next one is Shenma flavor.
They can always scan for "strange things". . .
(four) Go and take the risk!
You see this:
A formula, do you guess the Quark Scan King will scan the effect of the God Horse?
has the preview just now, and you may have become half an expert: First of all, quarks can at least scan the formula as a picture (just like I'm inserting screenshots of this formula into an article now); however, it has a more elegant choice - "extract" the formula like text.
Yes, I encountered a problem that is a hundred times more difficult than "text extraction": "format extraction".
Moreover, the "formula" is only one of all annoying formats at most. Let’s just look at a poster with mixed pictures and text typesettings, various fonts, and deformation designs that are curved into shrimps - all of them are formats!
format is the devil, anyway, I still remember the fear of being dominated by format in college. At that time, when I wrote a paper, the small problem of "indenting the first line" was often that one square was too left and one square was too right. Pureness could drive people crazy.
Quark This group of people needs to "subdue demons". In order to "die" the format on a picture, the masters went to other document teams to learn from the scriptures and embed a complete set of text format processing logic into the Scan King.
Things are not over yet. If you mix text and images together to type, the nightmare will be ten times more.
Let’s take scanning that poster as an example and ask soul questions one by one:
Problem 1: Where is the pattern and where is the text? The
poster layout is to let oneself go. Decoration, spots, words, letters and numbers of various colors are mixed together.
Many times, even people with vision 1.5 have to carefully distinguish which one is the "comma" and which one is the "decorative black dot in the background".
Problem 2: Is it a poster slanted or a pattern slanted?
There is an error in the human eye, and it is difficult to take a straight poster, and there will always be more or less deviant angles. At this time, the system should turn the poster into a regular order.
But the problem is, if there is an artistic word on the poster, the font is italic. So, how do you determine whether it is "the word "slanted because the poster is slanted" or "the word "slanted is not so slanted"? !
Problem 3: Is it a pattern or a flaw?
Some posters are big, and there will be creases after being expanded. So, how do you tell if a line is a pattern design on the poster, should it be preserved, or a useless crease that needs to be erased?
Problem 4: How to make up for the background?
poster scanned out with format, which means you can modify the words on the poster.
poster is no longer a PS file, all layers have been integrated. If you delete a word, a terrifying blank space will appear below it! What colors and patterns should be used to fill it?
Any problem in this is a difficult problem in the industry, but Huang Ruihua rolled up his sleeves and said that the labor and capital were not up to Akime Mountain. You really think I am a seller of tofu!
, such as "cream" and "pattern", may look the same from three meters away, but if you look at it in a magnified manner, you will definitely find out the difference between the two. The improvement that the algorithm needs to be made is "fine". But doing such a fine micro-operation means that cloud AI needs to calculate more data and consume more computing power.
, for example, "background supplement", is actually the most advanced AI algorithm that can supplement the gaps in the middle based on the surrounding information to make it fake and real. The problem is also "data quantity" and "computing power". Only by piled up machines in the cloud can you do this job well. The palace of
AI welcomes technology streams, but it is even more welcomes RMB players.
Huang Ruihua told me that in order to challenge the limit, they have gone crazy. In 2022, the AI computing power of the entire Quark Cloud has been upgraded to the latest generation of GPUs, and the algorithms running on it are basically upgraded and updated every month.
At this moment, although the scanning effect of the poster has not reached the perfect level in his heart, he can still shoot his chest compared to similar products around the world.
I tried it for you⇩⇩⇩
Our users have a high acceptance of typesetting. You can know from the membership payment situation that everyone is willing to pay for the good results.
Huang Ruihua, you're welcome.
tells us back to our story. Starting from 2022, the Quark Scan King suddenly entered an interesting spiral: the stranger the things that can be scanned, the more people use them; the more people use them, the more people will scan for stranger things.
As the saying goes, I don’t know what seafood is coming up when the waves are too big.
Recently, through various channels, the master began to see people taking playlists, some taking music scores, some taking texts from ethnic minorities, and some taking a bunch of source code.
Although they didn't say it, the user will definitely not do meaningless actions.
Huang Ruihua said.
So, the Scan Wang team began to think about some crazy things. For example, can a music playback engine be embedded so that the scanner can directly play the music when he sees the score? For example, can you add an emulator and run the code you see directly?
You see, from extracting text to extracting format, now it is "extracting everything". Scanning the pit is deeper than the Mariana Trench. . .
Of course, these are just a small number of examples. Various interesting functions are discussed every day, and may not be made in the end due to various factors. But the old master dares to discuss these things, provided that the core technology of scanning has reached a level that is sufficient to "support the work".
"Take the code as an example, it is extremely sensitive to format. In the same position, putting two spaces and three spaces will have different meanings. If the scan is not correct, the execution will definitely be wrong!" Huang Ruihua explained.
I feel a little emotional after talking about this.
users must not think so much when shooting these strange things. But in fact, far away, all their reliable and unreliable attempts were not blamed, nor were they ignored, but were thought about, remembered, and echoed.
If there is a "Internet spirit", I'm afraid this is an important part of the Internet spirit.
I once talked to Quark’s friends in 2020 (refer to "What are they saying when young people say quarks are really fragrant"
). At that time, I felt that the proportion of quark users was extremely high - those who were purely novice in life who did questions and searched questions all day long.
But nowadays, users who look at Quark Scan King are very different.
Among them may have a company's personnel manager who needs to regularly organize personnel information and back up the contract; there may be engineers who need to scan the drawings to add and modify the design; there may be "landlords" who need to scan an old rental contract, modify the information and print it again.
This picture contains a house rental delivery form, a company inquiry letter and a material plan.
These users have obviously walked out of the novice village of life and are fighting against monsters on their own battlefields.
This is a loving change in my opinion.
I tried to imagine how this happened: maybe people who used quarks when they were students, they entered a variety of jobs and were still used to relying on quarks to solve problems; maybe some "old people" were recommended by young people to "quark yyds" and then got into the trap. It may also be both.
But the reason why all this happened is probably because Quark did not disappoint them in a long period of time.
Most of the time, "maintaining trust" is more difficult than "building trust". Because you have to examine yourself all the time for a long time - not only your own good, but also your possible evil.
(Five) Brothers, believe me!
In the feedback received by Quark on a daily basis, some lawyers said that he used quark to scan the judicial documents, while some company finance officials said that she could scan business data.
You may feel a little uncomfortable when you see this.
is right, say it boldly - these data involve privacy. When using quark scanning king, will there be any problem with the data passing through the cloud? The big discussion that caused
did not first happen in Quark's user group, but within Quark's team.
In order to make the subsequent discussion based on facts, it must be emphasized here that no one in the Quark team can view the pictures uploaded by users, nor can they leak the data uploaded by users in any form. What restricts them is not only self-ethics, but also legal provisions.
But you have eaten a few bowls of noodles, it’s not your final say. The movement of the world cannot be achieved by swearing.
For example, there are many companies that state in the rules and regulations that internal data is prohibited from being uploaded to the Internet. Regardless of whether it will cause problems, uploading is not allowed anyway; there are also many special working environments that have signal blocking themselves. In this case, even if you are allowed to use it, it will not be able to use it.
After realizing these problems, Quark and other people decided to launch an "patch" urgently - scan offline.
Do you still remember the three-step scanning we mentioned at the beginning? 1. Local cropping; 2. Cloud processing; 3. Send back the local export file.
The so-called offline scanning is to stuff all the 2nd step (removing flaws, removing handwriting, text extraction, format extraction, etc.) into your mobile phone. The data does not come out of the mobile phone from beginning to end, so there will naturally be no privacy leakage.
, such as "not out of the phone", is verified - you can turn off the network, unplug the phone card, enter the signal shielding room, whatever, as long as the phone has electricity, the Quark Scan King can work.
However, this goal will cause a huge technical problem: the
algorithm team has just tried their best to maximize the computing power in the cloud, and to turn the algorithm to the latest, so that the scanning effect can be achieved to the extreme. Now you hold your phone and tell him, forget your cute and sexy little clouds, this small box is the home of the algorithm. . .
However, the irritable master Huang Ruihua did not overturn the table. Instead, he was very supportive.
He supports the idea of assigning a small team from a team that is not rich in staff and specializing in "offline algorithms".
puts an algorithm into a mobile phone, which is as difficult as putting an elephant in a refrigerator. Strictly speaking, it is more difficult than putting an elephant in a refrigerator, because cloud algorithms are 100 times larger than those in a mobile phone. There is nothing in the world that can still maintain its original functions after compressing 100 times - this means that most algorithms must be rewritten.
The master raced day and night. Finally, in August 2020, "offline scanning" was launched, using 1% model size to achieve a level of 90% in the cloud.
From one perspective, the offline scanning level is amazing, but from another perspective, it is a "very poor cost-effective" feature.
For example, after it goes online, there are not many calls - because in most cases, the user scans no sensitive information and will not turn on offline mode.
For example, sending a dedicated team to continuously maintain offline algorithms is very uneconomical - it is equivalent to a car running two engines at the same time, which is huge.
, and there is another very uncomfortable problem: when most users are not paying attention to privacy risks, actively launching offline scans will make them feel weird and risking to get angry. Just like I suddenly said to you: Do you guess the zipper of my pants is pulled up? Haha, it's up!
But Yoon Yibin said that they don't look at the problem like this.
"Whether a function is good or not, it cannot be measured purely by data. To be more polite, data can be lie, and sometimes you must trust your own head more," he said.
"But there will always be pressure on cost, right?" I asked.
"Of course there is pressure. Pressure is pressure, and the right things are right." He said, "Besides, the right things will sooner or later bring data returns! Don't worry, we won't lose in the long run."
tells the truth. When every young man grows up, life will teach him a compulsory course sooner or later, that is: When more and more people rely on you, something called responsibility appears in you out of thin air. Responsibility is responsibility, and it will not become the responsibility of others because you do not want to be responsible.
This is both simple and difficult. It is so simple that everyone can understand it, but it is so difficult that few people can do it.
, there are only three or five things with high cost performance in total, and people with normal IQ are doing it. Isn’t it true that people are different from people, products and products that are doing more, but the right thing to do?
can't be said that quarks are already a role model, I mean the direction they are working hard is quite sexy.
(VI) We each have a miracle of our lives
Wang Huai is a person who is full of imagination about the future. Although what he is doing is just a quark client, he believes that there are metaphors about the future in the small cameras behind the phone.
We do a lot of different things based on the camera, but there is a main task behind this - "understand the world" .
For example, the earliest recognition of everything we did was to understand the items in the lens; the scanning king we are doing now is to understand the information on a plane.
If you look to the future, we can point the camera at everything. For example, if you have a street view in your field of view, the name of the building will be marked on the screen; for example, if you have a book in your field of view, the author's era book review and other information will be displayed, and English books can also be translated into Chinese in real time.
In my personal opinion, the ultimate form of camera-based tools should be an AR (Augmented Reality) assistant.
he said.
I found some clips of future AR applications on the Internet⇩⇩
I recalled his description, and a strange scene suddenly appeared in my mind:
I am a Wushao on the first day of work, so nervous that I was blind. I even forgot what my colleague next door called me. At this time, I opened the scanner on the AR glasses, the other party’s business card, the time and place we met, and the content we talked about all appeared on the small screen a few centimeters away from the eyeball. It was simply a savior of social fear.
Just as I was staring at the screen and carefully studying who the person on the other side was, I was surprised to find out who he was also looking at his screen squintedly, struggling to find out who I was. . .
The scene is equivalent to: a game novice, standing in the wilderness with trembling trembling, suddenly a full-level divine sword fell from the sky, picked it up and wanted to kill all the people, but found that all the opponents around him had also instantly added the same sword. . .
Nowadays, everyone can sit in the belly of the iron bird and fly to any place in the world. Everyone can talk to friends in the sky with a chip. You won’t be surprised. Because when everyone has a divine sword, everyone has no divine sword.
Technology allows us to make progress equally, equally so that no one makes a fuss.
If there is really anything magical, in my opinion, it is the Quark team and their users, these two groups of young people have been "companying" from afar.
When a young man is on the road, he has an elf ball in his pocket. Although the monsters and monsters are welcomed by the way, they are extremely sure that there are a group of obsessive-compulsive disorder, veteran drivers, and good buddies who are as funny and curious as themselves and are unwilling to give up easily. They rack their brains to help them improve the elf ball day and night, and they may feel a little comforted.
Believe me, with my understanding of this fucking world for more than 30 years, even this kind of slight comfort is a rare and precious emotion in life.
can grow older with a group of friends with friendship, become older, and go from fighting the five scumbags to "old fighting the five scumbags". Repeatedly complaining about those bad jokes that only you understand, I am afraid there is no need to seek far away. The ultimate taste of the world is this pure joy.
Although you and I will eventually grow old, there are always people who are young.
There is one thing we are extremely sure of: in the corner of the world, there will always be 18-year-old boys. They look back at the entrance of the novice village, then hold the hilt of the sword, and walk towards the wilderness resolutely.
Maybe at this time, the BGM will still play the old song: I don’t care where you are going, where I am going. We will each have a miracle for our lives in the future.
Flip Pokémon
First there is the Pokémon
and then introduce yourself. My name is History , and I am a science and technology reporter who loves the story. My daily life is to chat with all kinds of masters. If you want to be friends with me, can search for WeChat: shizhongmax.
Oh, by the way, if you like the article, please don’t be stingy with your “reading” or “sharing” . It would be a wonderful thing to give interesting souls a chance to meet.
Thx within Beijing
