New Training Data for Kamstrup FlowIQ2200 water meter

[lauer]

Using the dig-class11_1500_s2.tflite model. It had some problems between 1 and 7 for the last digits, where there is a small line around it.

see #3121

Files

labeled.zip

[lauer]

I just downloaded and activated dig-class11_1701_s2 to see if that one is doing it better.

[lauer]

Did not work, again it confused a 1, and said it was a 7.

[SybexX]

Have you already tried the model dig-class11_1800_s2.tflite? you can find it there https://github.com/jomjol/neural-network-digital-counter-readout

test_models.zip In the zip is dig-class11_1800_s2.tflite from the link above and dig-class11_1801_s2.tflite trains with your images.

[lauer]

Ohh. Thanks for the quick reply. I didn't see the 1800 for some reasons.

Would you like me to check with that one first before checking the 1801 model?

[haverland]

Images are added to model https://github.com/haverland/neural-network-digital-counter-readout/blob/master/dig-class11_1801_s2_q.tflite in #25

Alternative the model https://github.com/haverland/Tenth-of-step-of-a-meter-digit/blob/master/output/dig-class100-0172-s2-q.tflite also includes the images

[lauer]

what are the difference between the models? Also, what does the _q mean?

[haverland]

class11 can only digits and "not a digit". Normally all that you need.

class100 and cont- can for normal digits on meters between the digit (in transition) the tenth of steps (0.0-9.9). Is for lcds not that really needed, but the models have massive more images to learn (>23.000). So it could be a better recognition. It's to try out.

-q.tflite vs. .tflite: -q means quantized. The models are smaller and faster with a minimal less on accuracy.

[lauer]

I can confirm, that the class11_1801_s2_q solves all the issues with 7 and 1. So I guess this ticket can be closed? Or at least when it's merged in jomjol/neural-network-digital-counter-readout#25