6、 Pointer meter recognition based on deep learning key points

Pointer meter recognition based on deep learning key points

Tips ： There are too many schemes on the Internet, all of which are segmentation ideas , Good idea , But it takes too much time to produce data , According to statistics ： Marking a piece of segmented data requires 180s-360s; The key point annotation is controlled in 30s-90s.

---

---

Specific implementation process

Tips ： The algorithm is based on deeplabv3 Semantic segmentation model and openpose The attitude estimation model is changed into a multi task learning model , Including key point detection + The pointer splits two parallel tasks , It is the core of the whole algorithm .

1. yolov5 Meter detection
2. Key point detection and pointer detection of the dial
3. Meter correction and coordinate transformation
4. Fit the arc structure of the dial , And calculate the ratio
5. According to the ratio 、 The range calculates the reading

---

Tips ： The following is the main body of this article , The following cases can be used for reference

One 、 yolov5 Meter detection

Target detection is needless to say . There are two main tasks in this step , First of all , Detect the meter in the image ; second , And classify each meter , In order to configure the range in the dial according to the category .

Two 、 Key point detection and pointer detection of the dial

This part is the core of the whole algorithm , We need to integrate key point detection and segmentation into a multi task learning model . Key point detection reference openpose, Split network reference deeplabv3 Wait for the mainstream network . Compared with Baidu's Algorithm （ Divide the scale and pointer ） The task of data annotation is reduced 60% above , More convenient , Easy to optimize .
The following figure is Baidu's plan ：

Want to know my plan ？ Then you can imagine turning the scale into a key point , The principle of pointer segmentation is the same .

3、 ... and 、 Meter correction and coordinate transformation

Use perspective transformation to correct the tilted meter , And transform the coordinates .

3、 ... and 、 Fit the arc structure of the dial , And calculate the ratio

According to the information obtained from the corrected meter , And fit the dial structure

3、 ... and 、 According to the ratio 、 The range calculates the reading

Finally, according to the ratio 、 Calculate the specific reading of the measuring range , The figure below shows the percentage , There is no time to transform , Make do with this diagram .

summary

This scheme realizes ：
Environmental Science ：pytorch、python=3.7、c++
Date marking time ： This plan 60s VS Baidu solution 360s

Tips ： So you are also willing to use annotation 1 This picture needs 6 Minute plan ？
A reliable plan , Let's develop 、 Optimize 、 Deploy 、 It's easy to land .