“It’s like we’re buying things on Taobao, we’re seeing sellers’ shows, but it’s the buyer’s show. The biggest difference is the light environment.â€
"The image-receiving seller show and the buyer show are always separated by a light." Eyemore founder & CEO Zhu Jizhi said at the IF Innovation Conference 2018 that the world's first AI imaging chip was released, "We It is often seen that image recognition rates are high in various image recognition competitions. However, in actual scenes, such as face recognition, no manufacturer dares to provide recognition rate data because there are too many on-site problems."
The picture below is a PPT shared by Zhu Jizhi at the press conference. On the left, it is part of a nude photo of Leina, the model of the Playboy magazine. This photo has rich details, clear layers and clear edges. All the images related to image algorithms have been used as standards for 30 years; but the pictures actually taken at the scene are often the right ones and cannot be recognized at all.
“The distance between the seller’s show and the buyer’s show of image recognition is always separated by a light.â€
"It's like we are buying things on Taobao, we are seeing sellers' shows, but it's the buyer's show. The biggest difference is the problem of light environment." Zhu Jizhi said, taking photos for beautiful women, will Set a lot of lights. However, the light environment in real life is uncontrollable. When encountering low light, backlighting, and reflective conditions, the imaging effect is very poor, and the AI ​​algorithm cannot recognize it.
The end of the pixel era for people to see, the visual era for machines to see
To solve the problems of the seller's show and the buyer's show, it is necessary to rely on the evolution of the visual organ to put the vision on a system as a whole. The first is the eye, which is responsible for sensing at the front end, producing images; then the brain, responsible for cognition at the back end, which analyzes image vision; and, in addition, the third part—how does the brain control the eye, that is, how the two sides interact intelligently? Only the three parts of the brain, eyes, and brain-eye interaction are intelligent, and machine vision is intelligent. This also represents the three stages of artificial intelligence development in the industry: the evolution of the brain, the evolution of organs, and the evolution of brain and organ interactions.
Three kinds of intelligence
To understand the evolution of the imaging organ of the eye, we must first review the history of imaging technology. Imaging technology began in the film era in the United States in the 1930s. The representative manufacturer was Kodak. In the 1980s, the digital era, the industry moved to Japan. All the digital photos seen today are from the imaging architecture of Japan in the 1980s, including Sony. Nikon, Canon and other companies. However, in the AI ​​era, the industry's demand for images may change substantially: the image is no longer for people to see, but for the machine.
When you look at it, the natural focus is on pixels. Girls who like to take selfies will definitely care about how many pixels the camera is before and after the phone. However, when everyone started to brush their faces with iPhone X, it seems that not many people care about the pixel problem of the camera. Because common sense tells us that when people look at the world, there is no pixel concept.
The human eye is the result of long-term evolution of human beings. The most powerful thing is the ability to adapt to the environment. Under normal circumstances, seeing everything is clear, the colors are correct, there is no problem with the seller show and the buyer show. In contrast, the biggest gap between the machine and the human eye is that the adaptability is too poor, and to solve the problem of adapting to the environment, the machine can only use three resources: computing power, algorithms and data.
In order to solve the problems caused by various complicated light problems, Eye Engine uses a variety of new algorithms, and the computational complexity is more than 50 times that of digital imaging. By testing a large amount of scene data, the imaging engine can be automatically like human eyes. Adapt to various environments, eliminate the effects of various light environments, and output stable visual images.
From IoE to VoE, the new billion-dollar market was born
There are two types of visual technology: imaging and image processing. The front-end imaging technology is responsible for generating visual images, and the back-end image processing is responsible for analyzing, identifying, and processing visual images. In other words, imaging is equivalent to the human eye, and image processing is equivalent to the human brain.
At present, star companies in the field of artificial intelligence, including Shang Tang, Yu Shi, Horizon, Yun Cong, Yi Tu, Shen Jian, etc., are all unicorn companies based on image processing algorithms. In the past three years, driven by deep learning technology, image processing has achieved rapid development, but the front-end imaging technology is still at the level of 20 years ago, which has become a serious development of AI vision and commercial applications. The bottleneck is also the next battleground for the major AI companies.
"AI will drive a major revolution and subversion of imaging technology and industry from pixel to vision. China is the fastest place in the AI ​​vision industry. I believe that the third phase of this new imaging technology will be dominated by China. Zhu Jizhi said.
Compared with the digital age, the imaging of the AI ​​era has undergone substantial changes in many key aspects such as imaging architecture, algorithm model, evaluation criteria, and light adaptability. The traditional digital imaging technology architecture can no longer meet the needs of AI vision, facing The dilemma of being quickly eliminated. In the next five years, imaging technology is expected to complete the epoch-making upgrade from “image†to “visionâ€. The visual imaging chip and AI processing chip become the core components of artificial intelligence, and the industrial upgrading demand derived from it will be in the next five years. It has spawned a new market of billions of dollars in the field of imaging.
From IoE to VoE, the new billion-dollar market was born
Configuring the vision center for the machine
Computing power, algorithms and data are integrated into a product, which is a chip, such as the Eyemore X42, the world's first AI vision dedicated imaging chip. The chip uses a new imaging engine architecture that integrates more than 20 new imaging algorithms with a sensitivity of up to 400,000, a single exposure dynamic range of more than 16 bits, and a maximum power consumption of less than 1.5W.
The world's first AI vision dedicated imaging chip Eyemore X42
The core imaging algorithms such as eyeMix and eyeNoise, which are completely developed independently, form the basis of X42. It abandons the traditional Japanese global imaging architecture and uses the sub-regional and layered Eyemore imaging engine architecture to solve the pain points of low light, backlighting, and reflection in visual imaging.
Eyemore Imaging Engine
"It’s really hard to do imaging chips. Eyesight Technology has been established for four years. Many people are asking me what I have done during this period? I can only smile and say that we are debugging images, debugging images, Constantly debugging the image. Because imaging is a subjective thing, we tested the scene above 500+, and it took four years before and after polishing to complete the world's first imaging chip for AI vision applications. "I can see that Zhu Jizhi is also invincible when he recalls the past."
The Eyemore X42 has only one mission, that is, imaging, which is to make the imaging engine eliminate the interference of the scene light in various complicated light environments, and output stable and reliable high-quality visual images to the AI ​​vision algorithm, especially under weak light. Beyond the human eye's visual imaging capabilities, it helps many AI company customers unlock more rich application scenarios. In order to improve the imaging performance of the chip, the developers even removed the standard video compression function. Zhu Jizhi said to the "Electronic Engineering Album" that this is like Intel's CPU integrated with the graphics card function, but Nvidia's dedicated GPU must be the mainstream of the future.
The future is a world of software-defined hardware, and Zhu Jizhi is also convinced of this. Therefore, in the X42 chip architecture, all the underlying imaging functions and various algorithms can be called. Unlike the "black box" attribute of traditional imaging products, the X42 chip is a "white box" that provides complete development tools and supports development interface APIs for various platforms including Windows, Linux, Android, and iOS. The purpose of this is to hope that all visual algorithm engineers can accurately control the imaging effect without having to know any hardware, thus improving the efficiency and accuracy of the AI ​​visual analysis algorithm.
Design-in 500 companies completed in three years
However, if a brand new chip comes out, who will believe you? Who dares to use? how to use?
Prior to founding Eye Engine, Zhu Jizhi worked for the largest chip distribution company in China for eight years and was responsible for promoting various types of chips. He knows the rules of the chip industry is that customers must prepare a series of solutions before using a brand new chip. The first is to have a development tool kit, so that customers can learn and research first; when the project is clear, there must be product modules to help customers to quickly productize; when the product is sold in batches and fully verified, the chip will be used directly; If the amount is large, IP authorization is also required; if the customer has special requirements, deep customization is also required. This complete process is the legendary Design-in. At present, Eyesight's AI vision product lifecycle full stack imaging solution is ready, and all customers can start using it instantly.
Eye Engine AI Vision Product Lifecycle Full Stack Imaging Solution
As we all know, the chip industry has its own fixed cycle. To be an original chip, from technology development to large-scale mature application in the market, it usually takes nearly seven years, that is, the "3+2+2" mode: three years of development, two years of promotion, two years of maturity. According to this rule, Eye Engine, which was established in 2014, will enter the scale promotion period after 20 years of development.
"3+2+2" mode for chip development
In Zhu Jizhi's three-year marketing strategy for the company, Eye Engine will focus on four market application directions: 1. Visual imaging for autonomous driving; 2. AI imaging for smart phones; 3. High-end intelligent security based on face recognition. 4, industrial visual imaging including military and medical. By the end of 2020, Eye Engine has become the global leader in emerging AI vision imaging technology by completing more than 500 AI vision customer design-ins and occupying more than 50% of the AI ​​vision imaging market. At the same time, Eye Engine will build a complete imaging ecosystem around visual imaging technology, and cooperate with all aspects of the AI ​​visual industry chain, including co-construction laboratories, strategic cooperation, joint development, and technology licensing to promote AI vision. The development will lay the foundation for the next stage to push "Eyemore Inside" to the hundreds of billions of visual applications.
“3D Structure Photoelectric Business Scanner Product†is a practical case shared by Zhu Jizhi on the spot. A manufacturer of "deep camera" used two imaging modules for e-commerce to scan 3D models of products, one for structured light and one for color images. After using the time-sharing scheme of Eye Engine, the problem was solved with only one imaging module. Then, through the interactive interface API, the efficiency and accuracy of the AI ​​vision algorithm have been greatly improved, which was previously unimaginable.
Eyemore Imaging Engine Application
Some thoughts on AI, vision and chip
â— The third kind of intelligence
The so-called "third kind of intelligence" actually refers to the relationship between AI and vision. What AI does is the brain, and imaging does it. There is a question here: How does the brain control the eyes? The traditional technical approach is to define a communication control interface, but this can be very complicated in visual applications. For example, the human eye has a characteristic, that is, "focus on one look." In general, the imaging of the human eye is very focused, only seeing the things of concern, others are vague. When the AI ​​algorithm solves the problem of "what to look at", the front-end imaging has a goal, and all the resources can be allocated to the object of interest, so as to "what to call". This kind of imaging according to the needs of AI can solve many problems that could not be solved before. Since the first half of the year, Eye Engine has focused on the development of interactive interfaces between the brain and the eyes, and hopes to cooperate with more AI algorithm companies to jointly promote the "third intelligence" of brain and eye interaction.
â— From one big to three countries
Zhu Jizhi said that in the past, mainstream processors have integrated image functions, including imaging and image processing, but the location is not important. This is the Intel mode; now, visual processing becomes the core, and the previously integrated visual part will be split. Come out and become a chip alone, this is the Nvidia mode, and other AI chips are also this idea. Similarly, the integrated imaging capabilities are not enough and will be isolated from the SoC processor, which is what the Eyemore imaging chip is doing. As the saying goes, "The world thing, the long-term must be combined, the long-term must be divided." Previously, the chip industry was a CPU monopoly. Now, because of visual reasons, it has become a three-point world. In other words, the visual impact of the chip industry landscape.
In the AI ​​era, the chip industry will change from a monopoly to a three-nation
â— Decentralized AI vision product industry chain
Blockchain is the most recent concept of fire, and its core idea is decentralization. Similarly, in the AI ​​industry chain, the decentralization process is also staged. Zhu Jizhi pointed out that the most important of the traditional hardware products is the CPU processor. The operating system runs on the CPU. Whoever masters this entry will become the center. For example, Intel, Qualcomm and MTK are the centers. But in AI products, the AI ​​algorithm and data run on the AI ​​brain chip including the GPU, and the CPU will no longer be the center.
In the era of centralization, the CPU will continue to integrate various functions and eventually become the turn-key SoC mode. When the chip integrates all the functions, the products made will become indifferent, and finally the only remaining manufacturing and sales capabilities. In the AI ​​era, even with the same chip, different products will have great differences and greater market value due to different algorithms and data. This is the biggest AI product generated after decentralization of the chip. value.
Decentralized AI vision product industry chain
The integrated power transducer can measure power parameters in three-phase power grid
with high accuracy, and has extended functions such as communication interface and analog output.
Integrated Power Transducer,3P4W Reactive Power Transducer,3P4W Sfere Reactive Power Transducer,3P4W Power Transducer
Jiangsu Sfere Electric Co., Ltd , https://www.elecnova-global.com