From the visual AI application case, what are the bottlenecks of AI landing?

The application scenarios of visual artificial intelligence (AI) are very rich and their commercialization value is huge. 40% of global AI companies are concentrated in the field of visual AI. At present, there are many companies engaged in visual AI in China, such as Shangtang Technology, Kuangshi Technology, Hongsoft Technology, Yuncong Technology, and Yitu Technology. They basically provide AI algorithms.

At present, the world's largest application scenarios for technology output are smart consumption, smart manufacturing and smart cars, as well as security and finance. Many projects have already landed in these application scenarios. What problems did these projects encounter in the process of landing, and where are the main bottlenecks? Has it been resolved now?

At a recent "Machine Vision Product and Technology Exchange" small salon, some machine vision companies shared the implementation of their specific projects and the problems they encountered.

Magnetic tile defect detection application case

Qian Xiang is an associate professor from the School of Advanced Manufacturing, Tsinghua University Shenzhen Graduate School. He is mainly engaged in the research of machine vision. They are currently engineering some scientific research results.


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

They started their operations at the end of last year and set up a Zhihan Equipment Technology Co., Ltd., which specializes in engineering research results of their graduate schools in accordance with the industrial application market demand. In order to avoid the red sea competition market such as glass cover and mobile phone appearance inspection, they chose a more subdivided field in the industrial scene application-magnetic tile defect detection equipment.

Talking about the reasons for choosing this segment, Associate Professor Qian Xiang boils down to three major reasons:

The first is that the market is large enough, as long as the machines that can "turn" need motors, whether it is electric toys for children, appliances, air conditioners, automobiles, or even space shuttles, motors need magnetic motors. There are usually two magnetic tiles in a motor, and some are four or more. According to statistics, China's current shipments of magnetic tiles account for about 70% of the world's total. It can be said that it is a trillion-level market, and it is still rising.

Second, the field of magnetic tile detection is still blank, and there are not many companies applying machine vision to this field. "There are reasons for fewer people, because for most people, this field is difficult and there are many pits." Qian Xiang said. For example, some magnetic tiles are black, some are bright, all sizes are different, the surface shape is different, some are chamfered, some are not, and the tile radian is different, there is no certain standard.

Third, the current status of magnetic tile inspection is mainly based on manual inspection. There are many problems with manual inspection. For example, because magnetic tiles generate a lot of dust during the production process, the inspectors will inevitably inhale too much dust during work, which makes it easy Causes pneumoconiosis; It is very hurting to stare at these magnetic tiles often, so every year to a year and a half or so, you need to change a batch of workers. Also the detection efficiency is very low.

From the perspective of the demand area, the areas with high demand for magnetic tile detection are currently concentrated in the Yangtze River Delta, the Pearl River Delta, Sichuan and Jiangxi.

After more than half a year's practice, currently the company where Associate Professor Qian Xiang is located has been in Guangdong and the Yangtze River Delta region, with 3 or 4 units sold.

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

Magnetic tile defect detection equipment shared by Associate Professor Qian Xiang
Figure 2: Magnetic tile defect detection equipment shared by Associate Professor Qian Xiang.

For example, he took a custom-made inspection device for a magnetic tile manufacturer in Jiangmen. For example, this customer needs to detect 12 faces. "We used 4 stations to flip the product and take pictures, and we used our deep learning algorithm to process the product. Dimension measurement. "Qian Xiang said.

Speaking of the experience of doing this project, he said with emotion, "Speaking of AI may be very exciting, but compared with natural scene recognition and natural language processing, all the industry left to AI is hard bones. We always feel that there is a good one Algorithms, adding a bunch of GPUs, CPUs, or NPUs can solve all problems, but this is not the case. AI algorithms alone are not enough. In the end, AI may not be the most important thing. The neck, the lighting, imaging, and electromechanics. "

"For visual AI, it also includes vision hardware." Qian Xiang concluded that when thinking about magnetic tile detection at the beginning, it was also very simple, that is, to detect whether the product was cracked, chipped, under wear, and air bubbles. It ’s just a matter of taking a picture. It ’s better to take a comparison, but in the end I found that taking pictures is not clear, because the magnetic tiles use different materials with different colors, some are black, they absorb light, some are bright, they reflect light ... … These must be taken into account when developing equipment. "It can be said that in addition to algorithms for machine vision, the hardware requirements are also very high. In the end, we spent a long time on hardware selection."

Also the electromechanical part is also very important. If the flowers are jittery during the operation of the machine, it will be difficult to obtain a clear image. Of course, this problem can be solved from the algorithm side, and it can also be solved from the mechanical side.

Regarding the imaging problem, he said, "In the beginning, we all thought it was simple, and it was enough to take a picture. In the end, we found that the picture was always unsatisfactory. Later, we had to improve our imaging system."

Connector defect detection scheme

Regarding the question mentioned by Associate Professor Qian Xiang, Yu Renchong from Aqiu Technology is deeply touched, because they are a company with deep learning, 3D vision and robotics as the core, using artificial intelligence in the field of industrial robots and automation. At present, products such as artificial intelligence industrial vision algorithm platform software AIDI and robot 3D visual sorting system solutions have been introduced for industrial complex visual inspection and disorderly sorting. The company is currently located in Beijing, Suzhou Kunshan and Shenzhen. Among them, Beijing is the R & D center, and Suzhou Kunshan and Shenzhen are the product and business centers.

The current problems in industrial applications mainly include: the uneven sample data amount of different categories, the uneven positive and negative sample data amount of the same category, the large change in defect size, and the inability to distinguish defect categories by simple image features.

In Yu Renchong's opinion, in addition to the four problems faced by industrial applications, the database is not easy to obtain. "Only after the customer believes in you can you get the data. After getting the data, you need an experienced engineer on site In cooperation with QC, we only know how to label the data. If the labeling is wrong, it will also cause interference. "

He also mentioned a problem. If the detection equipment has high accuracy and can detect all the problems, customers will be happy, but they will not use this equipment, because when all defective products are removed, the product failure rate will inevitably rise. High, which makes it difficult to ship new products.

In other words, if the AI products are to be engineered and implemented, they must get the customer's on-site cooperation. Because there are too many things to fine-tune, the debugging and training time of a project is actually quite long. He took the case of machine vision testing at the connector USB Type-C terminal as an example. For this project, there are about 30 indicators to be tested, and each indicator requires time to train.

Achu Technology's application case in the field of connectors
Figure 3: Achu Technology's application case in the connector field (1).

Achu Technology's application case in the field of connectors
Figure 4: Achu Technology's application case in the field of connectors (2).

Without the customer's on-site cooperation, AI products will be difficult to land, and engineering will be difficult to land. There are too many fine-tuning things here, and they are debugged on site every day. I think our industry is an energy-consuming industry, because it's adjusting and training every day, and the time is quite long.


Figure 5: Comparison between traditional vision manufacturer detection scheme and AI deep learning detection scheme.

In addition to the application of connector defect detection, Aqiu's AI algorithms and products are also used in industrial fields such as notebook computer case appearance defect detection, metal processing part appearance defect detection, blade tool defect detection, and other agricultural fields such as dried strawberry classification. There are applications.

Some smart IoT industry solutions

Beijing Kuangshi Technology Co., Ltd. was established in 2011. It is an industry IoT solution provider with artificial intelligence technology as its core. It provides artificial intelligence algorithms and solutions to industry users to build intelligent IoT systems. The products of Vision Technology mainly include face recognition technology, image recognition technology, smart video cloud products, smart sensor products, etc. The products are mainly used in finance, mobile phones, security, logistics, retail and other fields.

Wu Qiuyu, a senior technical expert in South China, shared some despised technology solutions for the intelligent IoT industry, including community real estate traffic management solutions, new retail, education, and logistics IoT solutions.

Despising Technology's Community Real Estate Traffic Management Solution
Figure 6: Community Real Estate Traffic Management Solution for Despised Technology.

Wu Qiuyu pointed out that in community real estate traffic management, the most important thing is to manage three categories of people and manage the four categories of people.


Figure 7: Wu Qiuyu is introducing Guangsui Technology's solution in the new retail industry.

He acknowledged that the current progress of new retail is not very smooth, but Desperate Technology has still done a lot of work. The main goal of its new retail solution is to open the data channels of people, goods and markets, and help intelligent upgrade of stores. Its technology mainly includes face payment, customer portrait and trajectory analysis.

People refer to intelligent understanding of members, to realize the experience of thousands of people, and to open the cashier payment to improve the efficiency of checkout; goods, refer to the intelligent selection of supply chain, improve the conversion of single store inventory, and automation of warehousing services, saving warehousing manpower Time; and field refers to the comprehensive capture of store data, providing real-time data feedback, and integration of store operations, real-time feedback, and real-time management.

He specifically pointed out that education is a key project of contempt technology this year. And since last year, they have begun to polish products in education.

Wu Qiuyu took the attendance system in the classroom as an example. The indoor time attendance system of Despise Technology is non-sense. You only need to install a camera with a PTZ and zoom in the middle of the classroom to scan the students in the classroom without dead ends. There are also functions for biometric identification and behavior analysis.

After the student finishes the class, all attendance information will be output to the host. Honestly, this behavior analysis is a bit "perverted". It can analyze the seven behaviors and seven expressions of a person, such as standing, raising hands, playing with a mobile phone, where the eyes are watching, etc. can be counted. All these behavioral analysis can be realized by a camera.

Although the author has graduated, thinking about such an attendance system scares a cold sweat. Although the technology is good, I still hope not to apply this technology to the classroom.

Educational solutions that despise technology
Figure 8: Educational solutions that despise technology.

Despising technology not only does college solutions, but also educational solutions for general schools, such as dormitory management systems. The traditional dormitory management system may also be a gate. Now, it is possible to add a face recognition system to the gate. According to Wu Qiuyu, such a dormitory management system has begun to be deployed in cooperation with Shenzhen University and Southern University of Science and Technology.

There is also a logistics manufacturing solution. Muangshi Technology has developed a robot network operating system--Muangshi River Map. Detective Technology and Cainiao built a smart warehouse with 500 robots in Tianjin. It is said that during the Double Eleven last year, it helped Cainiao improve its efficiency by 50%.


Figure 9: Logistics manufacturing solutions of Muang Technology.

But so far, Hetu has not reached the level of full commercial use, and is only doing strategic cooperation with large enterprises. The functions completed are also limited to automatic cargo classification, cargo forwarding and transportation.

3D visual artificial intelligence solution

AI 3D sensing technology solution provider Aobi Zhongguang also shared some of their application cases in visual artificial intelligence, and several interesting solutions.

Peng Xunlu, deputy general manager of Orbi Zhongguang, believes that Orbi Zhongguang is not an AI company, but an company that empowers AI companies. The key technology is 3D sensing technology.

Figure 10: Key technologies of Obilight.
Figure 10: Key technologies of Obilight.

He introduced the solutions of Aobi Zhongguang in the field of transportation, including subway face brakes and in-vehicle monitoring. According to Peng Xunlu, the subway face-washing brake has been deployed in Jinan Metro, Guangzhou Metro is about to be deployed, and Shenzhen Metro may be deployed next year.

In-vehicle monitoring mainly includes driver status detection, in-vehicle dangerous behavior monitoring, and bus passenger flow analysis. At present, Aobi Zhongguang is installing 3D cameras on 2,000 buses in Baoan District to count the passenger flow of the buses.

Orbi Zhongguang's Intelligent Transportation Solution
Figure 11: Aobi Zhongguang's Intelligent Transportation Solution.

Orbi Zhongguang's face brush payment solution
Figure 12: Orbi Zhongguang's face brush payment solution.

Now many self-service vending machines, some tea shops, retail stores, etc. can pay by face. Most of the front-end cameras of its face-brush payment system are made by Obilight. Peng Xunlu revealed that last year Ant Financial invested 200 million US dollars in Obi Zhongguang. This 200 million US dollars was actually not a financial investment but a strategic investment. Ant Financial and Aobi Zhongguang jointly established a company called Alio, which specializes in serving Ali Ecology. Its core business is Alipay face-brush payment equipment.

While cooperating with Ali, Aobi Zhongguang is also cooperating with UnionPay to develop UnionPay's face payment business.

Orbi Zhongguang Intelligent Robot Solution
Figure 13: Orbi Zhongguang Intelligent Robot Solution.

Intelligent robot, this is the earliest business of Aobi Zhongguang. According to Peng Xunlu, 80% to 90% of enterprises in smart robots have used Aobi Zhongguang's 3D cameras, mainly for obstacle avoidance and indoor navigation. "Everyone who uses a sweeping robot may have such a feeling. If there are shoes, cables, or obstacles lower than the sweeping robot, the sweeping robot will not actually recognize it and will directly hit it. The 3D camera can recognize some very small items, such as cables, animal feces, etc. "

He said that in the future, in terms of cleaning robots, there will be a complete set of solutions that can solve the problem of cleaning the whole house without requiring human intervention.

Orbi Zhongguang 3D Anthropometric Solution
Figure 14: Aobi Zhongguang 3D anthropometric solution.

In terms of 3D anthropometrics, Aobi Zhongguang has created a 3D fitting room for clothing brands, which can help customers place orders on their own and achieve clothing customization without visiting the store. He acknowledged that the device currently only supports products such as shirts and suits.

Peng Xunlu said that there is an experience store in Shenzhen CoCo Park, where there are 16 cameras in the 3D fitting room. Customers only need to stand there for 2 seconds, they can measure 28 data, and generate a complete body shape report, clothing Brands can use this report to help customers customize clothing.

Aobi Zhongguang's smart livestock solution in cooperation with foreign customers
Fig. 15: Smart animal husbandry solutions that Aobi Zhongguang cooperates with foreign customers.

Another interesting application case is smart livestock. Orbi Zhongguang not only does face recognition, but also pig face recognition, and establishes electronic files for pigs.

Peng Xunlu said with a smile, "Don't do not know, after we talked to the animal husbandry manufacturer, we found that