"The front-end imaging technology of computer vision is behind the huge demand of at least 10 billion orders of magnitude!" At the 2018 AWE scene recently concluded, Zhu Jizhi, CEO of Eye Engine Technology, was so excited to have reached such a conclusion. Zhu Jizhi's tone of speech is very certain, because their eyemoreX42 chip listed less than two months, has been singing all the way, showing off. The industry’s view on eyemoreX42 is that this world’s first fully self-developed and officially released AI visual imaging chip will “lead the AI ​​machine into the visual 2.0 eraâ€. Since this chip is currently unique in the domestic and even global fields, we can't help but chat with him a bit more. 2018 will be the first year of AI landing In recent years, the fieryness of artificial intelligence is unquestionable. It has been reported to the government twice that it has been written into the government work report, and the various AI companies that have sprung up after the rain show all these signs of a new era of artificial intelligence. arrival. Similarly, in many AI technologies and applications, computer vision is the biggest entry point and the most promising area. After all, in all the information we have obtained, the proportion of visual information has reached more than 80%. Therefore, this will be an unlimited market direction. The application areas of computer vision cover almost all industries that we know, autopilot, financial risk control/transactions, security, new retail, smartphones, robots... In the field of domestic computer vision, Shang Tang and Hao Shi are the two largest unicorns, and the financing competition they staged last year attracted the attention of countless people in the industry. First, in July 2017, Shang Tang Technology B raised 400 million yuan in rounds. US Dollars, and then on October 31st, Contempt Technology Face++ announced that it had won US$460 million in Series C financing. Shang Tang subsequently reported that it had received an investment of 1.5 billion yuan from Ali. Needless to say, we can feel the popularity of computer vision in the AI ​​world only from this series of record-breaking financing amounts. According to the relevant research report, the global computer vision market will reach 5 billion U.S. dollars in 2018, and by 2020, the market size of China's computer vision will increase to 72.5 billion U.S. dollars, and the future is extremely promising. “2016 is the concept year of AI, 2017 is the year of demo, and 2018 is the first year of the year.†Zhu Jizhi believes that in the next three years, AI will soon realize more extensive applications. "Although many technologies in the whole industry still have some problems, the overall structure has already come out," said Zhu Jizhi. Deeply cultivating the front-end imaging field, Vision 2.0 will bring qualitative changes to the industry ecosystem It is well known that currently, the technology companies in the AI ​​vision field in China almost all use traditional cameras, and all focus on the back-end processing of images. This has led to an embarrassing phenomenon: The vast majority of AI's algorithms are good enough, but the front-end imaging technology is not too hard, resulting in the machine's "eye" can not adapt to changes in the external light when the actual landing, the recognition of the poor effect of the embarrassing situation. Take self-driving cars as an example. Under unusual circumstances such as vehicles passing through tunnels, headlights on the opposite side of the car, night traffic lights blocked/disturbed by road lights, dark nights, dark fingers, etc., often due to poor recognition. A traffic accident occurred; In the same way, the security field is also unrecognizable due to black and white infrared images and backlit faces, and there are cases in industrial inspections that cannot be detected due to highlights and reflections. “Under complex light conditions, the SNR of the image captured by AI machines will be greatly affected. This is the biggest reason why AI vision products are difficult to land in the market.†Zhu Jizhi believes that the key to solving this problem thoroughly is to do Good front end imaging for AI vision products. To overcome the AI ​​machine's image recognition under complex lighting, we must first solve the image acquisition and front-end processing. At present, there are three mainstream solutions in the industry. The first is the laser radar we know well. In the case of poor light, the image and three-dimensional information of an object can be obtained by actively emitting laser light and using light reflection. The advantage of laser radar is very obvious. It can obtain extremely high angle, distance and speed resolution, and it also has good anti-interference ability. However, due to the huge size and expensive price (all types of 32-, 42-, and 64-line models, the prices are often tens of thousands, tens of thousands, and even millions, although they continue to decrease, they are still unfriendly). It is being marginalized. As Mr. Musk said, "Laser is like a crutch." This metaphor is appropriate. When visual abilities are not available, it is necessary to rely on a walking stick such as a laser radar. However, holding a crutch is always unsuccessful. The remaining two options are based on camera and imaging processing. The first is the array computing camera technology. The principle is somewhat similar to the compound eye of a multitudinous insect such as dragonfly and fly. The array of several, dozens or even hundreds of cameras captures a dynamic image with megapixel or more pixels, which is AI. The machine provides visual support. At present, this technology is still in the research and development stage, limited by the constraints of the volume and supply chain, and its products are still in the early stage of landing. The third option is eyeeyeX42, an eye-catching front-end imaging engine for AI. According to Zhu Jizhi, the way this eye does it, like human eyes, is to make people's eyes better - to solve the AI ​​machine's ability to automatically adapt to light in all kinds of light. “We firmly believe that the eyes of AI machines should be the same as ours in the future. After that, their eyesight will not be worse than human eyes and even stronger than human eyes. This is the original intention of our business, and we must fundamentally solve the vision problems of AI. ." Eye Eye EnginemoreX42 chip, visual 2.0 era AI machine imaging engine Regarding AI Vision, Musk once put forward the concept of “all-weather passive optical image recognitionâ€, which is to solve complex light, including accurate identification under low light, back light, and reflective light. This is just a problem that AI machines have to solve. Eye Engine Technology calls this vision that automatically adapts to light as "Vision 2.0." On the afternoon of January 19, 2018, at the 2018 geek park innovation conference, AO Vision Technology, the domestic AI visual imaging chip technology company, officially released the "eyemoreX42" chip. It is reported that eyemoreX42 is the world's first fully self-developed and officially released AI visual imaging chip. The data shows that the eyemoreX42 imaging engine chip, which has 20 times more computing power than traditional ISPs, employs more than 20 new imaging algorithms and integrates complex ray data over 500 different scenes. The eyemoreX42 chip has three features: first it is an independent imaging chip. Currently, there are imaging functions in various cameras, smart phones, and cameras, but they are all integrated in the main chip and can only be called integrated imaging. The eyemoreX42 whole chip only focuses on one thing, that is imaging. This is a bit like Intel's CPU has integrated graphics capabilities, but we know that only a dedicated GPU like NVIdia will be the mainstream of the future. The second feature, the eyemoreX42, abandons the traditional ISP imaging architecture and uses a new imaging engine architecture to solve imaging problems under complex lighting. The traditional ISP, in terms of architecture, cannot solve the complex light problem perfectly. The third feature is that eyemoreX42 provides a rich API interface, allowing the algorithm engineer to do back-end image recognition, and it is very convenient to control the imaging process. How powerful is the performance of eyemoreX42? Maybe the demonstration of the “difficult recognition in low light/dark light environment†demonstrated on the day of its launch will be a good illustration of the problem. "If you ask what the so-called glimmer is going to be, to what extent? It's very simple. We have a basic standard, which is to compare with the human eye. We just want to surpass the human eye," Zhu Jizhi said. There are about 125 million rods and cones in the retina of the human eye, and they act as photosensors. Among them, the rod cells sense the light and darkness of the light, while the cones are responsible for sensing the color. When the brightness of the light reaches a certain level (light or dark), the cones stop working and switch to the rod cells. At this moment, the human eye can only feel the gray level of black and white and temporarily lose it. The ability to perceive color. "What we are doing now is eight times the size of the human eye's 18 DBs. In that case, we can accurately restore colors when the human eye cannot see the colors and can only see the outlines." People's ability to perceive the color of the world is limited. Although in theory, human eyes can distinguish up to 12 million colors, it is actually far lower than this figure. But is it possible for the machine to distinguish between one million, ten million, or even more colors? This is entirely possible. It is hard to imagine what a machine can do beyond the human eye. Compared with low-dimensional vision, high-dimensional vision has unparalleled advantages. For example, when we see a flower, we see only white, but the machine sees it is colorful, because there may be 100 kinds of white. This is the core capability of machine progress and can see more information, so it can give more accurate feedback. Under AI Substantialization, Visual Chips Will Have 10 Billion Orders of Requirement in 5 Years With the further development of AI technology and the continuous expansion of applications, various machines/equipment driven by AI chips continue to emerge, making AI more and more show a trend of materialization. “Because of the wide application of artificial intelligence, a great technology ecology has been formed. In this huge ecological ocean, the new species of AI machines has begun to evolve, and these AI machines will quickly enter our reality. World." Zhu Jizhi said. The biggest feature of the AI ​​machine that distinguishes it from the general machine is that it is not an ordinary tool. It has its own brain. We can think of autopilot, robots, including process inspection equipment, smart security cameras/access controls/locks, etc., as an AI machine. On the other hand, from the information era to the current AI era, the core of competition has evolved from the competition of processing and manufacturing technology/hardware equipment performance to the competition of algorithms and algorithms. As a result, as the algorithm and the carrying capacity of the chip, it has become the biggest competition focus in the modern era. The AI ​​machines need more and more powerful chips, namely various AI chips. "The new species of AI machines is driven by various chips. In the era of algorithm and computing power, the number of chips needed by an AI machine, that is, the density of chips, will show an order of magnitude increase." Take Bitcoin, which has suddenly become popular, for example, its mining equipment—mining machine—is the big demand for chips. In an ordinary mining machine, there will be tens to hundreds of processor chips, which was impossible to imagine in the past. In addition, based on the Autonomous Vehicle under the Internet of Everything scenario, its vision system, route planning system, interior temperature regulation, external communication, etc., each individual processing unit needs at least one chip; the same There are also various robots, various drones, security monitoring, unmanned supermarkets... Zhu Jizhi believes that there will be hundreds of thousands of requests for various types of chips in the future of an AI machine. Driven by the internal cause of Vision 2.0, the era of large-scale application of various types of AI chips has already come. The new species of AI machines will open up a huge new market for AI chips. The demand for visual organs from AI machines is also huge, and the supply of AI visual chips will also increase accordingly. In the automotive field, a self-driving car will be equipped with 10 visual cameras; in the field of security, all surveillance cameras are faced with an upgrade of face recognition; in the field of industrial equipment, the detection of products through visual means , will become the standard for each production line, after each production line will be configured with more than 10 smart cameras; in unmanned retail, such as Amazongo is shifting the focus of attention from the label of the goods to the consumption of each customer It is customary that each of its unmanned retail outlets is equipped with more than 100 cameras. According to incomplete predictions, within the next five years, all kinds of AI machines will bring about 10 billion orders of magnitude of visual equipment demand, and the demand for visual chips will exceed this figure. "This will be an opportunity for all AI chip entrepreneurs." Guangzhou Bolei Electronic Technology Co., Ltd. , https://www.nzpal.com