Shulou(Shulou.com)11/24 Report--

In the past two years, rumors of machine vision companies known as the "four Little Dragons of AI", such as Shantang, Kuangshi, Yuntong, Yitu, capital market performance and layoffs, may make the public feel that machine vision is not a good business, and there is no hope in this field.

This is obviously not the case. On the one hand, computer vision (CV) is still one of the areas with the highest proportion of artificial intelligence technology applications, and related applications are the first choice for digitization and intelligence in various industries. Industrial quality inspection, patrol inspection, logistics robots and face recognition based on machine vision are landing more and more. There is no reason why AI is very good, but CV is not. On the other hand, the entire machine vision market is still growing. According to Forbes, the value of the computer vision technology market is expected to reach $48 billion by the end of 2022 and may become a source of continuous innovation and breakthroughs.

Obviously, it's not that the machine vision business is no longer sexy, but that the head machine vision company can't tell the story well.

However, in the face of geopolitical crackdown, the slowdown of domestic real estate infrastructure tourism, and the reality of capital markets at home and abroad, it is somewhat heartless to ridicule the dilemma of the "four Little Dragons of AI" with sarcastic words.

Of course, the real voices from industry users and developers also make it impossible for us to frivolously say things like "see the rainbow after wind and rain" and "keep the clouds open and look forward to the future." Because the reality is likely to be that when the economic environment recovers, some companies will still be able to recover, while others may sink.

Where can machine vision companies go when times change? Do you really see the changes of industrial AI? If you want to survive, what is the urgent need to carry out the work? This is what we want to discuss.

Is machine vision still a good business? The poor performance of machine vision enterprises represented by the four Little Dragons of AI is considered to be the disillusionment of the commercialization of machine vision and even AI. After all, these companies have an absolute lead in technical capabilities, business models and market share. If even they are difficult to make a profit, doesn't it prove that AI unicorns, which are good at algorithmic skills, are really "unmatched"?

Let's leave some of the media's habit of "reporting the bad news but not the good news" and take a look at the wider industrial world.

In the past few years, machine vision, as the most mature and widely used capability of AI technology, is actively embraced by the industry. Authentication in mobile phone applications, temperature monitoring at airport stations, active identification in the field of security, intelligent quality inspection in the industrial field, traffic flow and vehicle identification in the field of traffic. Perhaps the first AI ability that most people perceive is machine vision.

Today, it is almost a consensus to use machine vision as the entry point of transformation to promote industry / enterprise intelligence.

There is a complaint in the industry: "AI + industry can not escape the rut of machine vision." Is this really a flaw? We have interviewed many industrial enterprises, and most people have proposed that the first thing for their own enterprises to land is the application of machine vision here for quality inspection and inspection. A technician from the development of mechanical science told us: this is the current situation of most AI or deep learning landing, machine vision landing point to promote AI technology, is still a good strategy.

Even in sophisticated aerospace, machine vision is the first choice for some researchers to try deep learning. We once talked to an AI developer in the aerospace field about why many aerospace AI applications are image-like, and the other person said that the educational community had also discussed this issue. First, a relatively complete system has been formed in the field of image processing, image classification and target detection have been used a lot, and good results can be obtained quickly. Second, deep learning is data-driven, the aerospace field itself has accumulated a large number of optical image data, other load data can also be introduced into the AI model, but the effect is not as significant as CV. Third, from the perspective of the popularization of space knowledge, the image is very intuitive for ordinary people, and each picture taken by the Mars probe or the lunar probe can attract more public attention, and the social significance of taking the lead in introducing CV is a little greater. Therefore, at present, whether it is the rover or other spacecraft, there are many people in combination with CV technology to solve all kinds of problems.

It is safe to say that machine vision is an indispensable part of digital infrastructure. As people's image data in various fields such as work, entertainment, life and so on become more and more abundant, complex and sharp, the ability of analysis, processing and reasoning is also more and more, which is bound to push machine vision to a larger market.

Or think about it in reverse: if CV is really unimportant and there is no market, will it still become the target of "blockade" and suppression by neighbors on the other side?

Therefore, machine vision must be worth vigorously developing and betting on. Then there are new problems, the revenue and business potential of the "four Little Dragons of AI", which started with machine vision, are not optimistic, losses, breakers, layoffs and layoffs. Shangtang Technology said directly in its prospectus: there is still a possibility that it will not be profitable in the future. Does CV really have a future?

Interested readers may have read some analytical articles about the four Little Dragons of AI, attributing their poor performance to excessive investment in technological research and development, poor industrialization capability, encirclement and suppression of science and technology enterprises and traditional security enterprises, and so on. These factors exist, of course, but it is a dilemma for CV enterprises to change. Isn't it high for security enterprises to turn to machine vision? What if we reduce the R & D investment and lose the original technological advantage? There are barriers in the industry, which are aimed at all pure algorithm CV enterprises, so we can't have no way to lay down and quit collectively.

Since you can't lie flat, you have to think about how to turn the salted fish over. Then, the voice from industry customers and developers may be able to bring some "counter-common sense" thinking.

The first point of "anti-common sense": CV enterprise R & D investment is not too large, but not enough "AI four Little Dragons" is recognized to invest a lot in technology research and development, but with the decline in performance, this is regarded as a drag rather than a moat. There is a saying, "our chief investment officer thinks that what Shangtang can do, other companies can also do. Although Shangtang technology is more advanced, it may be three months to half a year apart."

Isn't technology a barrier? Big mistake, technology is definitely a huge or even the biggest barrier. Because at present, the number one worry hindering the large-scale application of machine vision in the industry is the actual performance of CV technology, which is not up to expectations.

There are probably a lot of common problems in CV landing industry after combing.

1. The effect of replacement is not obvious. You may think that brushing the rankings in AI competitions is a very powerful CV, but it should be noted that increasing from 85 points to 90 points is a breakthrough in academia, but in industry, it may take more than 95 points to be applied. Moreover, the competition is just a group of deep learning systems to compete behind closed doors, in the real world, AI is faced with very mature algorithms that have been used by enterprises for a long time.

Ma, a teacher from Zhengzhou University, was invited to develop defect detection for a non-woven enterprise in Jiangsu Province. He told us that there is a very mature detection algorithm Halcon in the industrial field. After long-term accumulation and iteration, the algorithm has been very stable, and the amount of calculation is small, and there is no need to label data and adjust parameters. Compared with the cost of transformation and application, machine vision needs to find very attractive and differentiated application scenarios and functions.

two。 The performance of real application is declining. The "top students" who perform well in the laboratory become "poor students" as soon as they land in the production environment. Teacher Ma told us that the nonwovens are spreading out at a speed of 36 kilometers per hour, which requires machine vision algorithms to accurately draw the location of defects under high-speed movement. the technical challenge is very great. At this time, the recognition speed of the traditional algorithm is faster than that of the CV algorithm because of its small amount of computation and mature algorithm. And the scene environment is not standard and controllable, sometimes the light is reflected through the glass in the morning, and the light spot is produced on the non-woven fabric, which will affect the detection accuracy of the CV system, and the false alarm rate is very high. Teacher Ma and his team went through all kinds of debugging to make the accuracy reach a stable value. For enterprise customers, they would like to introduce mature technical products, which must be made by CV enterprises or R & D personnel to tell customers that they can really improve efficiency, and some enterprises are willing to try.

3. The limitations of the application scenario. Quality inspection and patrol inspection may be the key scenarios of CV, but for factories and other enterprises, they may only be one of the links of the production process, so more enterprises may choose to introduce AI in the form of "packaging" solutions in the intelligent process, and machine vision may be just one of them. A technology company told us that when doing an intelligent transformation of the production line for a factory, the total project amount is 50 million, of which the vision may only be 50 to 1 million. And more industrial enterprises such as metallurgy, rail transit, manufacturing and so on pay more attention to CV quality inspection, these areas are also the Red Sea of CV enterprises.

The interviewer said bluntly that a deputy director of the Ministry of Industry and Information Technology once led a team and took ten AI enterprises to each company to look for business cooperation opportunities. It was found that some petrochemical companies proposed to use machine vision inspection instead of manual inspection, but the tower of the other party was very high, so it was impossible for the robot to climb the stairs and maintain high stability. Finally, as soon as the AI companies evaluated, they found that they could not do it. There are still many similar situations, and many enterprises are really concerned about the pain points. At present, CV technology can not give enough substantial help.

You must have found that there is still a big gap between the real world and the industrial world in the CV algorithm, which is easy to crush humans in laboratory and AI competitions. Even for some key industries and landing scenarios that are eager for AI and suitable for CV, such as quality inspection and patrol inspection, there are still some unsolved technical problems.

From this point of view, how can the commercialization of CV enter the exponential growth rapidly, and the head CV enterprises have failed to bring a bright solution to the problems that customers in these industries are really concerned about.

AI technology, including CV, does not have so many gimmicks and tuyere. The prerequisite for successful application is the maturity and cost performance of the technology itself. The road to R & D is long and has far-reaching significance, and it is also worth adhering to and eventually opening the gap.

The second point of "anti-common sense": the poor revenue of CV enterprises is not because they do not understand the industry. Many people think that AI companies with pure algorithms tend to have strong technology but cannot be transformed when doing the B-end market because they do not understand the industry and the scene. This is part of the reason, but it is not the most fundamental reason.

The difficulty of deep integration of AI and industry is well known. The four little dragons of AI basically have their own deep-rooted industries and scenes, and are actively trying to transform from pure machine vision to a more comprehensive AI solution service provider. For example, the Internet of things solution, which combines end-to-side hardware with AI algorithm, focuses on solving computing problems according to the map, Shangtang's "1 (basic research) + 1 (industry combination) + X (industry partner)" model, cloud from focusing on core areas such as finance, travel, etc. In other words, no one wants and works harder to understand the industry than they do.

A deeper reason may be that you know, but you don't fully understand, and you may never fully understand.

First, in terms of cognition, pure algorithm companies and the physical industry have natural barriers.

At present, the word "intelligence" proposed by intelligent manufacturing in the industrial field is different from that in the AI field in terms of concept and application details. For example, the AI world tends to aim at a certain CV task, build a model, and learn data characteristics, so that the model has the ability to solve specific problems, that is, pure algorithm ability. However, the "intelligence" desired by industry and even many industries is actually partial to the entity, such as the flexible transformation of production lines, the interconnection of manufacturing equipment, and digital twin technology. In this case, when CV enterprises begin to enter the industry, there will be the phenomenon of "talking with chickens and ducks".

Second, in practice, the high labor cost of pure algorithm companies does not meet the economic benefits of mass customization.

The digital transformation of enterprises requires "thousands of people and thousands of faces", not only in different industries and fields, but also among different companies in the same industry. The scene of CV application is very fragmented and requires a high degree of personality customization. This leads to a contradiction that CV enterprises have to rely on a large number of algorithm engineers with high labor costs to solve the needs of all industries, which is obviously impractical and expensive.

Take the industry as an example, different processes, manufacturing requirements and specialties will result in completely different mechanism models needed in the machine vision system, even among different products in the same field, and the complexity and process threshold are also relatively high. The person in charge of an industrial enterprise once told us that, for example, in the defect inspection of metallurgical steel coil production, the steel coil is divided into cold rolling and hot rolling, which is equivalent to four problems. although they are all machine vision quality inspection, the mechanism problems to be solved by AI are completely different, so this enterprise or a scene can not achieve the generalization of the model. The customized development of one-to-one model will lead to the increase of landing cost and implementation cycle.

The brain team has visited many intelligent projects, and many of the details are unimaginable in the lab. A pig farm hopes to use AI recognition to monitor pig temperature. As a result, the detection effect is not good after the system is online. A check found that the pig skin is too thick and the temperature identification is not correct. There is also a waterworks, all replaced with smart cameras, but usually it is not open, because the high-precision video images captured by the cameras are too large, the factory's network is not upgraded synchronously, and the speed of uploading analysis is very slow. Some enterprises directly produce other colors of cloth on the production line without notifying the technical staff, which can't be adjusted to the CV system all of a sudden.

An AI developer told us that an AI hardware product for piston ring testing was developed for a factory, in which tens of thousands of solder joints were soldered manually, the welding process was broadcast live on Douyin at that time, and 300000 screws were screwed. No software company will hire an algorithm engineer to screw, but precisely because no one does it, many of the systems produced will not conform to the actual situation of the factory, he said. or the factory will not be convinced of the plan put forward by the enterprise. And it is precisely because he will communicate very directly with the factory and have done the most basic and boring manual work, so the proposed AI plan is generally quite convincing.

In public reports, the "four Little Dragons of AI" have said that a large part of R & D spending is on talent recruitment. On the one hand, the scale of income is not proportional to the cost of manpower, and the revenue situation is not good; on the other hand, the intelligent demand of fragmentation also determines that it is impossible for enterprises to fully cover and reuse on a large scale in their efforts to expand manpower.

In the words of that AI developer, there may be millions of processing plants in China that need to use AI. Different products in a factory may need different models. It is impossible to rely on engineers of some algorithm companies to complete. There may be a large number of front-line workers. As long as you learn to use AI, you can apply the more mature CV algorithm to the production line.

In other words, the intelligence of thousands of industries inevitably requires a variety of CV applications and models. When machine vision technology changes from generalization and simplification to diversification, personalization and scene, it is also doomed to head or large-scale CV enterprises, it is impossible to train all the models and eat all the markets on their own.

The third point of "anti-common sense": CV enterprises seem to be besieged by the market, but in fact, the commercialization of peer CV enterprises is not ideal, which is indeed impacted by the transformation of traditional security giants into AI and the entry of some AI technology giants into CV. After the demise of these giants with data, channel, technology, ecological and other advantages, it seems that some common CV market segments begin to "roll" prices, so that machine vision companies such as the "four Little Dragons of AI" are very hurt.

But on the other hand, is the CV market a special industry monopolized by unicorns?

Industrial intelligence must be accomplished by many AI developers, ISV service providers and CV enterprises. In CV enterprises, there are not only giants, but also a large number of small and medium-sized enterprises. For these enterprises, the empowerment from the open platform is precisely the hope of survival and development.

Some entrepreneurs in the field of vision tell us that they will not deal with Party An in the name of a start-up company at all, and they may not even get the money back. Their company is as a cooperative supplier of some large enterprises, large enterprises for customers to do the overall system integration solutions, this CV start-up to solve one of the small visual problems. If a small enterprise does the project alone, it may encounter a variety of situations during the project acceptance.

"from our implementation experience, now the AI visual surface is so widely spread that there is no good solution to many problems, there is still a certain distance from the requirements of Party A. For example, the customer's quality rate is close to 100%. In the acceptance phase, the production line is required to run for a week. The error rate cannot exceed three times in a week, and no acceptance will be given to you if more than three times. As for the AI model, it may be easy to adjust from 80% to 95%, but it is very difficult to get from 95% to 96% or 97%, so collecting money is a very difficult thing. "

In addition to the difficulties in the business model, small and medium-sized CV enterprises are also faced with difficulties in recruiting people and setting up barriers. As mentioned earlier, the versatile CV track is already very crowded, price war is obvious, and transparency is very high, many enterprise customers are very clear about the cost of AI system. "AI four Little Dragons" is still uncomfortable, small and medium-sized micro enterprises are naturally more difficult to set up barriers in the Red Sea.

However, on the other hand, enterprise customers do not care what technical methods are used by technical service providers to solve problems, computer vision based on deep learning can also be used, and traditional machine vision can also, as long as it can meet the needs of the application. and the cost is acceptable. This requires the overall cooperation from the basic hardware, algorithms, and then to the industrial level, which is also an opportunity for small and medium-sized CV enterprises, that is, they can integrate diversified software and hardware and intelligent capabilities to meet the needs of buyers. At this time, the technology giants'AI open source open ecology, the introduction of all kinds of machine vision models, algorithm market ecology, and so on, make them the objects to be selected and integrated, and their technologies and products are re-packaged into seed products, thus expanding market share.

Some ISV service providers said that when developing the solution, they chose Kang Nishi's camera, which comes with an algorithm library for industrial vision, which can be bought back to directly develop more targeted products, and then buy a brand of encryption dog when deployed, so there is absolutely no need and motivation to buy a pure algorithm company's algorithm.

Once upon a time, algorithm companies such as the "four Little Dragons of AI" did become the darling of the times through the marketization of algorithms, but when the baton of industrial intelligence was handed over to more diverse developers and service providers, compared with the competition for excellent developer resources by technology giants, large CV enterprises were indeed slow to respond and had limited advantages. While they are worried about share prices, more individual developers and small and micro business developers are already writing code and screwing in factories and fields.

The disappearing machine vision company has shared so many truths from developers and enterprises that we have to give advice to CV enterprises. First of all, machine vision will occupy a very important share of the digital economy, this market has never stopped development, there are a large number of scenes need image processing, image recognition, target detection and other capabilities, CV can develop a lot of industries. However, we may see fewer and fewer machine vision companies with pure algorithms as their core competencies in the future.

Technically, CV needs to be combined with other AI technologies to solve the complex problems that the industry is concerned about but has not yet solved.

As mentioned earlier, there are still a large number of scenarios that can not be solved by AI in the industry, even in the areas of quality inspection and patrol inspection, where AI is relatively widely used, there are a large number of blank detail scenes, which are of industrial value. But CV cannot achieve a breakthrough as a single technology. For example, the combination of CV and hardware, instead of people to some towers, minefields, mountain areas to work, requires the comprehensive ability of machine vision, robots, chips and so on. For example, the "four Little Dragons of AI" are all in the field of intelligent medicine, and some doctors have said that pathological images contain a lot of image information, and after AI extracts these information, they can conduct more in-depth research with genomics and proteomics, rather than just such simple segmentation or classification.

In terms of business model, for some small and medium-sized enterprises, through the major algorithm platforms and the development of out-of-the-box tools / interfaces provided by ecology, they will become integrated AI solution providers in the future, rather than focusing on machine vision capabilities. For companies with advantages in the CV track, integrate machine vision capabilities with the comprehensive technologies and capabilities needed by camera, processor, 5G, cloud and other industries, and support more developers and industry partners to create models with higher industry value and complexity. The reconstruction of business models and successful platform will broaden the depth and width of the machine vision industry.

Behind the disappearing machine vision company, there is the background of this transitional era: some people become the clouds of the past in the ebb and flow of the tide, and some people become the winners in the deep water.

This article comes from the official account of Wechat: brain polar body (ID:unity007), author: Tibetan fox

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.