Fast high-resolution image processing with just one NVIDIA GPU
Available in Cognex Blue, Red, and Green options

Cognex has released 'Cognex BD Suite', a deep learning-based machine vision software. It has improved on the defect detection in irregular patterns and pattern recognition with many variations that existing machine vision technology could not do. It announced that it will attack the domestic vision market through this.

Machine vision is a technology that recognizes and collects sophisticated patterns and features in images, and then performs precise guidance, inspection, measurement, and high-speed reading, and is necessary for increasing automation and efficiency in smart factories. During the manufacturing process, precise placement is important in the alignment stage. If accurate placement is not achieved, it can take a lot of time and money, which can increase the overall cost or the proportion of defective products.

The most important feature of machine vision is precision. It finds and confirms minute details and features down to micrometers that people cannot do. The camera captures, identifies, and measures precise parts such as line segments and edges at the front end, and corrects them if they deviate from the alignment position. This is why machine vision is necessary in automation.
Existing machine vision has shown limitations in detecting defects in irregular backgrounds, detecting complex patterns with many variations, and recognizing characters. Vision has had difficulty inspecting defective patterns that people can recognize at a glance. For the above reasons, Vietnam's manufacturing market currently has more than 500,000 people involved in inspection. Although it is efficient for overall automation, it has shown weaknesses in the detailed detection required for 'customized production' in smart factories.

Machine vision enables inspection by rules created through various signal processing. It is difficult to create rules that detect irregular and highly variable patterns that people can immediately recognize. Cognex Vidi Suite goes beyond rule-based and applies human learning to deep learning to enable pattern detection that only people can solve.

“By combining existing machine vision to increase automation and new product solutions, we can increase productivity, reduce defects, and improve efficiency,” said Min-Soo Kim, Cognex Vision Solutions Manager.

IBM's Watson beat everyone else in the famous quiz show Jeopardy Show in the US and won first place. Also, IBM's Deep Blue looks at the chessboard and makes the optimal move by looking at five moves ahead based on the current move. These are all rule-based systems and have surpassed people in many aspects.

One game that I thought I couldn't win was Go. I thought it was impossible to approach it in a rule-based way because the search space was almost infinite. However, Google's AlphaGo broke this. AlphaGo judges the Go board as an image and judges each point as three values: black, white, and empty space. AlphaGo's learning ability is trained with a huge amount of data called Go records. Here, artificial intelligence plays against each other to conduct reinforcement learning.
Data is important when learning a network. Facebook, Google, Amazon, etc. have big data. They create artificial intelligence and train it on a large amount of data to operate like a human. Deep learning has gradually developed due to abundant data, high-performance computing, and deep learning algorithms. However, there are several problems in utilizing this deep learning in the industry. Google and Facebook have servers to utilize a large amount of data. However, the industry does not have a large amount of data. In addition, it is difficult to apply deep learning centered on big data because they do not have servers.

In other words, it must be able to process even small amounts of data. Cognex Vidi Suite has solved this problem. Since it does not require a large amount of image processing at once, high-resolution image processing is possible with a single GPU that can process at high speed. A large number of servers are not required. In addition, it provides functions optimized for factory automation based on Vidi Suite Blue, Red, and Green tools. Unlike existing deep learning-based machine vision solutions, it can learn with image data sets of several tens to a maximum of a hundred sheets. Additionally, it is capable of optical effect processing, high-resolution color and thermal image recognition to perform inspections accurately.

Conventional machine vision only distinguishes between passes and fails and leaves the learning to deep learning, but Vidi is composed of blue, red, and green and processes various detailed features and options.

VidiBlue is used to find and localize multiple features in a single image. Even objects with complex features, such as severely deformed characters on a noisy background or multiple complex objects, can be located and identified by recognizing annotated images. To train the Blue tool, you just need to provide an image with the target features marked on it.

Bidi Red is used to detect anomalies and aesthetic defects. It learns the normal appearance of an object, including distinct but acceptable variations, and can identify numerous problems, including scratches on decorative surfaces, incomplete or improper assemblies, and the texture of fabrics.

Bidi Green can be used for object or whole scene classification. It distinguishes different classes based on a collection of labeled images, such as product identification based on packaging, classification of weld seams, and separation of acceptable and unacceptable defects. The green tool can be trained through images that are assigned and labeled according to different classes.

Cognex CEO Moon Eung-jin said, “With these three tools, even non-vision experts can easily design and use them flexibly. And since they can process small amounts of data, they can be applied to both individual consumers and industrial applications.” He continued, “Conventional deep learning has a long learning time, making it difficult to apply to industrial sites when the model changes, but Cognex is capable of fast learning because it uses less computing power and images.”