From the case of visual AI application, what is the bottleneck of AI landing?

Visual Artificial Intelligence (AI) has a rich application scenario and huge commercial value. Forty percent of AI enterprises in the world are concentrated in the field of visual AI. At present, there are many enterprises engaged in visual AI in China, such as Shangtang Technology, Ignorance Technology, Hongsoft Technology, Cloud Cong Technology, Itur Technology, etc. They basically provide AI algorithm.

At present, smart consumption, smart manufacturing and smart cars, as well as security and finance are the most widely used scenarios in the world. In these application scenarios, many projects have already landed. What problems did these projects encounter during landing? Where did the main bottleneck come from? Has it all been solved now?

At a recent "Machine Vision Products and Technology Exchange" small salon, some companies engaged in machine vision shared the landing situation of their specific projects and the problems they encountered.

Application case of magnetic tile defect detection

Qian Xiang is an associate professor from the Department of Advanced Manufacturing, Shenzhen Graduate School of Tsinghua University. He is mainly engaged in machine vision research. At present, they are engineering some scientific research achievements.



Figure 1: Associate Professor Qian Xiang is sharing the origin of Chinese equipment.

They took action at the end of last year and set up a Zhihan Equipment Technology Co., Ltd. to engineer the research results of their graduate school in the light of the market demand for industrial applications. In order to avoid the competition of glass cover and mobile phone appearance detection in the Red Sea market, they chose a more subdivided field - magnetic tile defect detection equipment in industrial scene application.

When it comes to the reasons for choosing this subdivision, Professor Qian Xiang concludes that there are three main reasons:

First, the market is big enough, because as long as the "turning" machine needs motors, whether it is children's toys, or household appliances, air conditioning, automobiles, or even space shuttles need motors, and motors need magnetic tile. There are usually two magnetic tiles in a motor, some of them are four or more. According to statistics, at present, China's magnetic tile shipments account for about 70% of the world's total, which can be said to be a trillion-level market, but also rising.

Secondly, the field of magnetic tile inspection is still blank, and there are not many enterprises that apply machine vision to this field. "There's a reason why there are so few people doing it, because for most people, this field is very difficult and there are many pits." Qian said. For example, some magnetic tiles are black, some are bright, the size is various, the surface shape is also different, some have chamfered, some have not, the radian of tile is also different, there is no certain standard.

Third, the current situation of magnetic tile testing mainly depends on manual testing, manual testing will have many problems, for example, because magnetic tile in the production process will produce a lot of dust, the inspectors will inevitably inhale too much dust in their work, which is easy to cause pneumoconiosis; and often stare at these magnetic tile, it is very eye-injuring, so every year to one and a half, it is necessary. We need to change a batch of workers. And the detection efficiency is also very low.

From the point of view of demand area, the areas with high demand for magnetic tile testing are mainly concentrated in Yangtze River Delta, Pearl River Delta, Sichuan and Jiangxi.

After more than half a year's practice, Professor Qian Xiang's company has already sold 3 or 4 sets of orders in Guangdong and Yangtze River Delta.

In the field of magnetic tile testing, "we have developed a set of such equipment in our company based on our existing technology, including optics and algorithms. We can customize a testing production line according to different shapes of magnetic tile." Qian Xiang said in sharing.

Magnetic tile defect detection equipment shared by associate professor Qian Xiang

Figure 2: The magnetic tile defect detection equipment shared by Professor Qian Xiang.

He gave an example of a magnetic tile manufacturer in Jiangmen who had customized testing equipment. The customer needed to test 12 surfaces. "We used 4 workstations to take flip photographs of the product, and we used our deep learning algorithm to measure the size of the product." Qian said.

Speaking of the project, he lamented, "It may be exciting to talk about AI, but compared with natural scene recognition and natural language processing, the industry leaves AI with a hard bone. We always think that a good algorithm, plus a bunch of GPUs, CPUs, or NPUs, can solve all the problems, but in fact it is not the case. AI algorithm alone is not enough. In the end, AI may not be the most important thing. Lighting, imaging and Mechatronics are the bottlenecks.

"For visual AI, it also includes visual hardware." Qian Xiang concluded that the initial magnetic tile testing, thinking is very simple, is to detect whether there are cracks, cracks, underwear, and gas bubbles in the product, just take a photo comparison, but ultimately found that photographs are not clear, because the use of different materials for magnetic tile color is different, some are black, will absorb light, some are bright, will be counter-productive. Light... These must be taken into account when developing equipment. "It can be said that besides the algorithm, machine vision also requires a high level of hardware. Finally, we spent a long time on hardware selection."

And the mechanical and electrical part is also very important. If the flowers shake badly in the process of mechanical operation, it will be difficult to get a clear image. Of course, this problem can be solved from the aspects of algorithm, as well as from the mechanical aspect.

As for imaging, he said, "At first we all thought it was easy, just take a picture. Finally, it was found that the photographic effect was always unsatisfactory, and then we had to improve it.

Please read the Chinese version for details.