most people have some misconceptions about ai vision chips, thinking that ai companies that do image processing chips have mastered all the vision technology, but not.” in china, the visual unicorn competition for visual processing technology has developed signiﬁcantly, and in recent years is also gradually applied to smart phones, security monitoring, automatic driving, medical imaging, intelligent manufacturing and other ﬁelds.
however, the visual processing chip cannot be separated from the acquisition of information and must rely on image sensors. on april 28, dr. wu nanjian, a researcher at the institute of semiconductors, chinese academy of sciences, told the daily economic news that the so-called visual chip is actually an on-chip chip with high-speed image acquisition and real-time image processing functions. t he chip is actually an integrated system chip with high-speed image acquisition and real-time image processing.
he and his team successfully developed a new vision chip and published a paper in 2011 , but up to now no company has realized the integrated chip of “image sensor vision processor”. the company has not yet achieved large-scale mass production of “image sensor vision processor” integrated chip.
this will deﬁnitely overturn the existing industry pattern. t he interview with wu was conducted during the ﬁrst y angtze river delta global science and t echnology innovation project focused roadshow, which was also the ﬁrst roadshow for the artiﬁcial intelligence vision system chip project. how to commercialize the visual system chip? when will it be accepted by the industry? wu nanjian said: “large enterprises do not dare to do to grab the rice bowl, then you can give the startups to do, this is a slow process of transformation.
the following is the daily economic news reporter (hereinafter referred to as nbd) interview with dr. wu nanjian.
nbd: artiﬁcial intelligence vision system chip and the existing vision chip what is the diﬀerence?
wu nanjian: artiﬁcial vision is divided into two parts, similar to the human eye and brain. t he human eye is a typical image sensor, able to take images and some noise removal and other primary image processing; human brain neuron network is a visual image processing system, with a very strong ability to parallel processing of the visual information taken.
at present, domestic and foreign research in the ﬁeld of artiﬁcial vision chip is mainly 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 towards the development of high resolution, wide dynamic range, high frame rate, high intelligence, wide wavelength range and three-dimensional imaging direction. artiﬁcial vision system chip can complete image acquisition, and primary (image ﬁltering), intermediate (feature extraction) and advanced (feature identiﬁcation and irregular processing) three image processing steps.
as the basic research on ai vision technology continues to deepen, the market pattern has developed into a relatively independent and interdependent industrial ecology. in the front end, sony is the image sensor market, production and technology leader,followed by samsung and howe t echnology also maintains a good competitive force; in the back end, mobileye and nvidia (nvidia) is the main provider of visual processing chip manufacturers, the ﬁeld of domestic companies such as horizon.
the artiﬁcial intelligence vision system chip is a high-speed cmos image sensor, parallel signal processing unit and output circuit integrated in a single chip to achieve real-time vision chip system, which is equivalent to a disruptive chip for the existing industry. t he integration of diﬀerent functional technologies on a single chip has many advantages, simply put the vision system chip has a higher match in terms of processing power, speed, power consumption and cost.
however, regardless of the current start-ups or large enterprises that already have a certain share in the market, they are either doing image sensors or back-end vision processors.
nbd: why have no companies chosen to try to integrate image sensing and processing on the same chip?
in fact, sony is doing this, as you can see in the 201 7 annual report, they have a team doing research on artiﬁcial vision systems, but not big. i have also exchanged chip design experience with them, sony is interested in this, but they have concerns, said to do integration may face an immeasurable situation.
not to mention that sony has put a lot of eﬀort in the ﬁeld of smart phones, only from the image sensor market competition pattern, there are sony, samsung and howie three enterprises, nce the visual chip integration, on the one hand, if the last two united, may shake sony’s position in the whole market; on the other hand, because the back-end (visual processor) enterprises to take away the rice bowl, the industry ecology will have a devastating the impact of the industry ecology. then do the visual processor companies are also the same idea.
but from my personal point of view, will the visual system chip become an inevitable trend? i think it will. like cell phones and cameras combined into a smartphone, the current technology has broken through the low ﬁll rate, low resolution and signal interference serious problems, the scientiﬁc research results and put on the market is only a matter of time. the most worrying aspect for large enterprises is that if a completely innovative enterprise does it, there is no such concern.
nbd: if the visual system chip is industrialized in the future, how big is the market space?
wu nanjian: on this point, we have done a projection. 2018, the market size of the image sensor in about 15 billion u.s. dollars, although 12 billion u.s. dollars in the ﬁeld of smart phones, but the future development of the four faster areas are security, defense, automotive, medical, by 2021 will welcome 4 billion u.s. dollars of market space, the annual growth rate is about 10% to 20%.
visual processor demand growth will be faster, the overall size of the market (including hardware, software, services) in the $ 17 billion to $ 18 billion, from the hardware alone also accounted for about $ 3 billion. if the visual system chip can cover a market size of $ 7 billion, the enterprise in this middle to get 1 % of the market size, its proﬁtability is already very large.
nbd: then, so that ai vision technology really from the laboratory to the application of landing, what are the entry barriers?
wu nanjian: at present, the products based on the technology has been used in some innovative enterprises, such as in the ﬁeld of industrial products automation inspection can be completely used vision system chip instead of manual inspection; in the ﬁeld of intelligent monitoring, the past need to install the visual processing chip in the camera with sensor technology, through the data structured, and then compressed to the data center to complete the complex way of data transmission and calculation, and then this structure may be destroyed.
it is important to understand that ics are capital-intensive, technology-intensive, and talent-intensive, so each of the thresholds is very demanding for companies. in terms of technology, after decades of research, at least the more diﬃcult problems are understood at this stage, and the core problems have been overcome in the laboratory because it integrates two kinds of chips: electrical and optical, which only a few teams in china can do.
but want to achieve real commercialization there are still many thresholds, ﬁrst of all, the problem of capital, has been relying on national projects to support the industrialization is unlikely, so the need to seek the support of social capital; secondly, the issue of talent, the main concern in the early stage of technical research and development of talent preparation, but later more need to join the engineering team and the market team to better enable corporate customers to understand the new vision system chip the advantages and practicality of the new vision system chip.
最先出现在kong tak electronic。