Let’s start with the first topic, what is the value of AI to industrial vision algorithms?
The traditional algorithm, I define it in terms of two key words, quantitative analysis and feature engineering. Feature engineering is counting different features. For later judgment, a bunch of if… else… is a feature of traditional algorithms. So AI, which you already know a lot about, is learning by sample. I think it’s a very essential thing.
From the functional point of view, industrial vision algorithm can be divided into image processing, positioning, detection, measurement, recognition; From the perspective of algorithm implementation technology is classification, recognition, measurement three categories. In essence, industrial vision algorithms will be, or are being, completely reconfigured by AI, although when it comes to measurement techniques, namely quantitative analysis techniques, traditional algorithms are still indispensable.
The value of AI reconstruction industrial vision algorithm is reflected in the following three aspects:
1. Increase dimension
AI solves some of our complex classification and recognition problems by raising dimensions. Including complex backgrounds, low contrast, flexible electronics, and some strong interference. These things used to be fine with traditional methods, but with AI, I think it can be better. This is a point that you can see with the naked eye. The second and third points may not be visible to the naked eye, but are actually more critical.
2. Simplify and generalize
One of the advantages of AI is that it can make extreme abstraction of algorithm problems, after which complex industrial vision problems will become simpler. Another is generalization. Many of the more complex algorithm problems in industrial vision, with two or three more general algorithm modules to train the data, the results come out, and this index is also very excellent.
3. Cut costs
When you hear this, it seems counter-intuitive that AI requires computational power. How can you still reduce the cost?
Let’s take an example that anyone who does traditional algorithms can understand. For example, geometric shape matching, which belongs to the whole machine vision, the traditional algorithm around such an algorithm, it needs to set a lot of parameters. In order to use it well, engineers need to understand the basic principle of geometric matching algorithm and the physical meaning of parameters, which requires professional background knowledge of image processing and high threshold requirements. If you don’t understand it well, it may not be the result you want, or it may not be a very precise effect. To do this, you need a background in image processing algorithms. So I used to do traditional graphics, that is, when I was in the previous employer, we took under the application engineers are master, this cost is very high.
However, if we use AI to do it, for example, if we only train three or five samples, or even one sample, then the accuracy and efficiency of the whole positioning can reach or even exceed the accuracy of the traditional algorithm. Of course, the overall robustness is definitely better than traditional algorithms. So the cost of use can be very low.