The development history of industrial visual inspection machine

It is an inevitable development trend that the detection of automatic detection machines replaces manual detection. The demand is becoming more and more diverse. It was only on the production line. Now there are various applications in various fields such as manufacturing, medical treatment, electronics, and storage. For example, a well-known domestic e-commerce company uses 3D to detect the package size and match the total parameters for internal tracking.

-2D to 3D: The industry is basically three-dimensional components, and two-dimensional imaging is after all a kind of pathological data collection of the actual situation in three-dimensional space. Therefore, various detection, measurement, robot guidance and other projects surrounding 3D are emerging in endlessly. This is similar to the situation of computer vision. There are too many companies for structured light, ToF, binocular and other technologies.

-The implementation of the system is greatly restricted by various factors: if you ask me what is important to do a machine vision inspection project, I will definitely say that I can get a high-quality picture, and just to get such a picture, you need to consider too much Too many, light source, lens selection, sensor selection, tempo considerations, installation layout, automation integration, environmental factors considerations, changes in workpiece status, etc., which part of the problem will affect your image quality, and if you don’t have enough Good pictures, no matter how powerful the algorithm is, and because the production line has yield and beat requirements, if the entire visual inspection system is not fast and reliable, your false detection rate will be very high, and this is what countless manufacturers need The problem faced. For example, I need to detect that there are several holes in a certain area of an aluminum housing. When the outer surface color of one batch given to you by your supplier is different or uneven from another batch, then wait What you have to do is to stop and re-set the parameters and re-verify the product. This is almost inevitable to happen 100% of the time.

-Algorithms are not as fast as hardware development: This is my feeling, and I think this is a major bottleneck at present. From a system perspective, hardware is developing towards a mobile embedded direction, while the current intelligence of software is far from enough. The deep learning neural network that is popular in the field of computer vision currently has very few successful applications in the field of machine vision. ViDi is one, Fanuc/Preferred Networks and Google’s robot crawling is one. This is still relatively simple. When the target object is changeable, the features are complex, and the number of samples is not enough, you have no chance to use deep learning at all, you still have to return Coming to the traditional old road, and considering the strict requirements of real-time, machine vision especially needs a new intelligent method that is commonly used in most application fields, whether it is innovation, improvement, or synthesis.

-To treat specific issues in detail: machine vision is still deeply related to specific application fields. Each application needs to choose dedicated hardware and software to match with it, and it requires special design, so no one solution can be applied to all Case.

-Company: There are a lot of integration agents in China, and there are too few leading technologies. Look at Lingyun and Daheng, you will know what the situation is. Recently, there is the old Dahai Kangwei for video surveillance, and it is also doing machine vision. . You can study Cognex and Keyence abroad. Also, no company will do everything in a machine vision system. The core is the algorithm and the integration of the entire hardware. There are specialized companies for lenses, specialized companies for light sources, and specialized companies for brackets. No company will Do it all by yourself.

Then I want to talk about machine vision and computer vision. Although they both process image data, there is a big difference between the two. One is more system-level application-oriented, and the other is focused on The algorithm is a combination of half theory and half application. But it is not without connection, such as the application of AR in industrial manufacturing, you can not tell which technology is to be classified.