Digit confusion between 1 and 7

[Mika-ncy]

The Problem

Hi !

There is an issue on my working system. The digit 1 on my counter is missread into a 7.

Si there any solution to learn the system not to confuse ?

Version

v15.3.0

Logfile

[2d21h12m47s] 2023-09-28T09:48:52 <ERR> [POSTPROC] main: Raw: 074748.0, Value: , Status: Rate too high - Read: 74748.0 - Pre: 14148.0 - Rate: 60600.0
[2d21h15m46s] 2023-09-28T09:51:52 <ERR> [POSTPROC] main: Raw: 074748.0, Value: , Status: Rate too high - Read: 74748.0 - Pre: 14148.0 - Rate: 60600.0
[2d21h18m46s] 2023-09-28T09:54:52 <ERR> [POSTPROC] main: Raw: 074748.0, Value: , Status: Rate too high - Read: 74748.0 - Pre: 14148.0 - Rate: 60600.0
[2d21h21m49s] 2023-09-28T09:57:55 <ERR> [POSTPROC] main: Raw: N44NNN.5, Value: , Status: Rate too high - Read: 44148.5 - Pre: 14148.0 - Rate: 30000.5
[2d21h24m44s] 2023-09-28T10:00:50 <ERR> [POSTPROC] main: Raw: 074748.0, Value: , Status: Rate too high - Read: 74748.0 - Pre: 14148.0 - Rate: 60600.0

Expected Behavior

No response

Screenshots

No response

Additional Context

No response

[caco3]

Please provide a screenshot of your device

[TopGun2010]

Hey - I got the same problem over here. Quite annoying. And yes, I see what the problem is - the lighting produces a reflection on the upper edge of the counter wheel. But I am not able to get the light at any other angle or whatever, then it might occour at a differenz number wheel and the problem is more to the right/left. I think this could be also solved by a better recognition - as a 1 looks differenz from a 7, the problem sometimes also evolves with other numbers e.g. a 2 is also recognized as a 7.

Using 15.7.0

Any chance for a better recognition?

[TopGun2010]

Tried everything before and again now with RC5 of 16.0.0, seems as if I really have to get rid of the reflections anyhow.

[caco3]

If the image does not contain the needed information, the ESP can not invent it. It is like on a normal camera, if you have noise, there is no way to remove it by software (without degrading the other image quality).
The right step really is to get rid of the reflections.

[TopGun2010]

I understand what you're saying, but I was curious why the 1 and 7, which do look different are not recognized as such. I mean, you can see the small thing of the 1 on the top, the 7 has an angeled stroke, the 1 is straight. Are you trying to cover different font styles? Are there styles where the 7 also got a straigt stroke like the 1? This is a thing which I cannot understand, cause when this is given the reflection wouldn't matter. But dont wanna have a discussion here at this point - I gotta get rid of the reflections somehow.

[caco3]

Keep in mind that there is only one model for all trained digits.
The neuronal network just outputs the best fit. This does how ever not mean that it is the right/expected result.

The model is a tradeoff between being generic (so it works for all kind of digits) and a good enough recognition.
The ideal ways would be to have a well trained model containing just those digits which you have. But this would be a lot of extra work.

[TopGun2010]

Okay and the tradeoff is as computing power is so limited I guess? And also of less training data? I'm not familiar with file sizes or how much more training data causes how much more computing power or much lager file sizes.
I just saw what was possible when some ppl of our Astronomy discord programmed a neuronal network to determine and eliminate color/light gradients in astro pictures. Which was quite impressive and became known in the astro community world wide in just a few months, but - admittedly it uses PC/MAC CPU power or even GPU/cuda :D But its more complex and pictures are more likely in a 26 MP range. So I was falsely thinking there would be more possible here. :)

[friedpa]

@SybexX The problem with the tool-tips/wiki is, that you and @caco3 spending a lot of time for explaining things over and over again. It seems, that a lot of users are simply lazy, why reading the tool-tips/ wiki, if I can get an answer without using any brain-power...
In other Git-hub forums there is a standard text send out pointing to the tool-tips/ wiki... just my fifty cents....

[caco3]

@friedpa Thanks for your feedback. Let us discuss this in a separate issue: #3444

I would be happy for your feedback and if possible also contribution.

[TopGun2010]

Just to defend myself or proove that I also read it, I think the problem why many ppl are writing here might be something different. Might be my understanding. For me it is not logical, why zooming in the problem would lead to another readout. Also higher gain and less LED power results in the same result in my understanding. Coming from photography I know when I got less light, but more sensor sensitivity/amplification (gain), the reflection stays the same relatively to the brightness of the picture, only the noise raises significantly, results in less quality signal and recognition.

For me it is not a faulty image you would be training with, it is something where it should be recognized as well, just other circumstances. And I bet many ppl got very similar problems which would be solved then. But your argument concerning limited computing power is more than an argument for me.

Don't get it wrong, love what you're doing, I just wanted to maybe get you into an "Apple-like" approach, to make it easy to use and kind of more fool proof.

Thanx for your time!

[caco3]

yeah but your reflection is smoother

That's because I added a Diffusor to reduce the reflections.

[TopGun2010]

yeah I tried that, too.. but cannot get it that smooth as it is in your pic. It still remains some kind of strong stroke. Will try more and harder.. Just so fiddly.

thanx guys!

[caco3]

Do you have access to a 3D-Printer?
I am designing a case for the ESP32-CAM which will contain a diffusor.

[TopGun2010]

Unfortunately not. I will try again by diffusing it, maybe I can do better this time.

[TopGun2010]

Happy New Year btw :)

[Ingo-Knito]

I have one :-) Maybe I can help?

To eliminate reflections - in my optinion - it is best to use external LEDs.
I've designed a housing for the krom schröder BK G16 - which has the same housing as the G10.
Here with two external LED

[caco3]

yes, that looks much better.
For my water and gas meters I work with the internal LED and the reflections are not disturbing (the gas meter has a diffusor made of white paper). But since so many people have issues I thought a 3D printed diffusor might be useful.

[TopGun2010]

Yeah, I know the best solution would be with external LED, but my initial point was, to make it as easy as it can be. With no soldering etc. But yeah, physics of reflections make it hard with the internal LED.