search

jomjol / AI-on-the-edge-device

Easy to use device for connecting "old" measuring units (water, power, gas, ...) to the digital world
https://jomjol.github.io/AI-on-the-edge-device-docs/
5.89k stars 641 forks source link

Save recognized numbers on error #2542

Open MarcFu opened 1 year ago

MarcFu commented 1 year ago

The Feature

When the result of a roundtrip is an error (e.g. negative rate), save the pictures of the recognised numbers along with the result number.

I want to know, what happens in vase of an error. Often the problem solves itself within the next rounds. Ti investigate, what is happening, this would definitely help.

caco3 commented 1 year ago

Such errors are usually due to the problem every neuronal network has, it picks the closest option. And since the models contain all kind of meters, it is always a compromise. This means, if eg. a digit gets close to 5.9, it could be read as 6.9. This even get worse as some meters do not show the digits very accurate, eg. the flip to early or to late. There is nothing we can do against this except the validations (max rate, neg rate) which are already implemented.

BTW, if you have a good SD-Card, nothing speaks against writing every round to the card, I do it since months without issues.

MarcFu commented 1 year ago

The implemented measure help a lot. But the question is, which digits get regularly misread.

In some cases I have the same digit multiple times, looking nearly identical but produce different output.

Maybe I have to try to train a specific model for my digits.

MarcFu commented 1 year ago

Here is an example: Screenshot_20230805_194620_Chrome

caco3 commented 1 year ago

It is odd, the almost identical 9 get once recognized as a 9 and also as 1...