??"Most people have a misconception about AI vision chips. They think that an AI company that makes image processing chips has all the vision technology. In China, the competition for visual unicorns has led to the rapid development of visual processing technology, which has been gradually applied in the fields of smart phones, security monitoring, autonomous driving, medical imaging and intelligent manufacturing in recent years.
??However, the vision processing chip is inseparable from the acquisition of information, must rely on the image sensor. Since artificial intelligence technology brings the opportunity of "decentralization" of chip market, various functional chips have emerged one after another. Wu nanjian, a researcher at the institute of semiconductors of the Chinese academy of sciences, told China Daily on April 28 that the so-called vision chip is actually an on-chip integrated system chip with high-speed image acquisition and real-time image processing functions.
??Wu nanjian is a researcher at the state key laboratory of semiconductor superlattice, institute of semiconductors, Chinese academy of sciences. In 2011, he and his team successfully developed a new vision chip and published a paper. However, up to now, no enterprise has realized the large-scale production of the integrated chip of "image sensor + vision processor".
??"It's going to disrupt the industry." The interview with wu was held during the first Yangtze river delta global science and innovation project centralized road show, which is also the first road show of ai vision system chip project. How is the vision system chip going to be commercialized? When will it be accepted by the industry? "If big companies are afraid to compete for jobs, they can leave it to start-ups, which is a slow process of transformation," says Mr Wu.
??The following is an exclusive interview with Dr Ng by NBD:
??NBD: how are ai vision system chips different from existing vision chips?
??Nj: artificial vision is divided into two parts, similar to the human eye and brain. The human eye is a typical image sensor which can capture the image and perform some primary image processing such as noise removal. The neural network of human brain is a visual image processing system, which has a very strong ability to process the visual information in parallel.
??At present, the research in the field of artificial vision chip at home and abroad mainly focuses on CMOS image sensor chip technology, parallel image processing technology and CMOS integration technology.
??In the field of CMOS image sensor, the current international technology level is developing towards the direction of high resolution, wide dynamic range, high frame rate, high intelligence, wide wavelength range and 3d imaging. The artificial vision system chip can complete three image processing steps: image acquisition, primary (image filtering), intermediate (feature extraction) and advanced (feature recognition and irregular processing).
??With the deepening of basic research on AI vision technology, it has developed into a relatively independent and interdependent industrial ecology from the perspective of market structure. On the front end, SONY is the market leader in image sensors, production and technology, followed by samsung and Howe technologies. In the back end, Mobileye and NVIDIA are major providers of vision processing chips, with companies in the field such as horizon in China.
??Artificial intelligence vision system chip integrates high-speed CMOS image sensor, parallel signal processing unit and output circuit into a single chip to realize real-time vision chip system, which is equivalent to a disruptive chip for the existing industry. Integrating technologies with different functions on a single chip has many advantages. Simply put, vision system chips have a high matching degree in processing power, speed, power consumption and cost.
??However, whether it is a startup or a large enterprise that has a certain share of the market, they are either making image sensors or back-end vision processors.
??NBD: why has no enterprise chosen to try to integrate image sensing and processing on the same chip?
??Nj: actually, SONY does that. In the 2017 annual report, you can see that they have a team working on artificial vision systems, but they haven't done that big. I also talked to them about chip design, which SONY was interested in, but they were concerned that doing integration would be immeasurable.
??Not to mention that SONY has invested a lot of efforts in the field of smart phones. Judging from the market competition pattern of image sensors, there are three companies, namely SONY, samsung and Howe. On the other hand, the loss of back-end (visual processor) enterprises will have a devastating impact on the industrial ecology. So do companies that make vision processors.
??But from my personal point of view, will vision chips become the inevitable trend? I think so. Just as the combination of mobile phone and camera makes the smart phone, the current technology has broken through the problems of low filling rate, low resolution and serious signal interference. It is only a matter of time before the scientific research results are converted into the market. As for the biggest concern of large enterprises, if it is a completely innovative enterprise to do, there is no such concern.
??NBD: if vision system chip realizes industrialization in the future, how big is its market space?
??Wu nanjian: we have made calculations about this. In 2018, the market size of image sensor is about 15 billion us dollars. Although 12 billion us dollars is in the field of smart phones, the four fields that will develop rapidly in the future are security, national defense, automobile and medical industry. By 2021, the market space will reach 4 billion us dollars, with an annual growth rate of about 10% ~ 20%.
??Demand for visual processors will grow even faster, with the overall market (hardware, software, services) now between $17 billion and $18 billion, or about $3 billion in hardware alone. If vision chips cover a $7 billion market, and a company gets 1% of that market, it's a lot of money.
??NBD: so, what are the barriers to entry for AI vision technology to truly move from the lab to the application floor?
??Wu nanjian: at present, the finished products based on this technology have been tried by some innovative enterprises. In the field of intelligent monitoring, visual processing chips used to be mounted on cameras with sensor technology to complete data transmission and calculation through the complex way of structuring and then compressing the data to the data center, which may break this structure in the future.
??It should be noted that integrated circuit has significant capital -, technology - and talent-intensive industry characteristics, so each threshold is very high requirements for enterprises. In terms of technology, after decades of research, at least the difficult problems have been understood at this stage, and the core problems have been overcome in the laboratory. This is because it integrates the two kinds of chips related to electricity and light, which only a few teams in China can do at present.
??However, there are still many barriers to achieve true commercialization. Firstly, there is the problem of capital. It is impossible to rely on national projects to support industrialization. Secondly, there is the problem of talents. In the early stage, it mainly focuses on the preparation of talents in technology research and development, but in the later stage, it needs the participation of engineering team and marketing team to better enable enterprise customers to understand the advantages and practicability of the new vision system chip.