Suggestions for improving my digital meter readings

[mt7479]

Hi everyone, this meter has been quite a challenge but I've managed to get it working with an error rate of 3 per day.
https://www.messtechnik.at/wordpress/produkte/warmezahler/kompakt-warmezahler/funk-wohnungswaermezaehler-modelle/elf2/

The display has no backlight and the digits are very dim, maybe the battery is almost depleted. The positing of the led is also quite tricky, if i try to aim it above or below the digits nothing is visible on the display at all. Aiming it on the left and sacrificing the last digit got me some usable results. The downside of this is that the digits to the right are a bit dark.

I currently have some false readings with a 5 recognized as 9.

So far i've tried all available digits models, I'm currently using dig-class11_1701_s2.tflite. Normal and greyscale mode but I always get some errors.

Heres my ROI Setup:

Whats also a bit puzzling, the seconds digit seems to be tight enough drawn, however in the recognition it picks
up a portion of the third digit. Is this some kind of aligment marker issue ?

Any ideas how to improve this would be greatly anticipated. I'm also thinking about training my own model.
Would the image quality be sufficient for this ?

Config:

[TakeImage]
;RawImagesLocation = /log/source
;RawImagesRetention = 15
WaitBeforeTakingPicture = 2
CamGainceiling = x8
CamQuality = 10
CamBrightness = 0
CamContrast = 2
CamSaturation = 2
CamSharpness = 0
CamAutoSharpness = true
CamSpecialEffect = grayscale
CamWbMode = auto
CamAwb = true
CamAwbGain = true
CamAec = true
CamAec2 = true
CamAeLevel = 2
CamAecValue = 600
CamAgc = true
CamAgcGain = 8
CamBpc = true
CamWpc = true
CamRawGma = true
CamLenc = true
CamHmirror = false
CamVflip = false
CamDcw = true
CamDenoise = 0
CamZoom = false
CamZoomOffsetX = 0
CamZoomOffsetY = 0
CamZoomSize = 0
LEDIntensity = 5
Demo = false

[Alignment]
InitialRotate = 173.0
SearchFieldX = 20
SearchFieldY = 20
AlignmentAlgo = default
/config/ref0.jpg 218 275
/config/ref1.jpg 521 261

[Digits]
Model = /config/dig-class11_1701_s2.tflite
CNNGoodThreshold = 0.5
;ROIImagesLocation = /log/digit
;ROIImagesRetention = 3
main.dig1 320 87 60 120 0
main.dig2 368 87 60 120 0
main.dig3 414 87 60 120 0
main.dig4 463 87 60 120 0
main.dig5 506 87 60 120 0

;[Analog]
Model = /config/ana-cont_1300_s2.tflite
;ROIImagesLocation = /log/analog
;ROIImagesRetention = 3
main.ana1 432 230 92 92 false
main.ana2 379 332 92 92 false
main.ana3 283 374 92 92 false
main.ana4 155 328 92 92 false

[PostProcessing]
main.DecimalShift = -1
;main.AnalogToDigitTransitionStart = 
main.ChangeRateThreshold = 2
PreValueUse = true
PreValueAgeStartup = 720
main.AllowNegativeRates = true
main.MaxRateValue = 2
;main.MaxRateType = AbsoluteChange
main.ExtendedResolution = false
main.IgnoreLeadingNaN = false
ErrorMessage = true
main.CheckDigitIncreaseConsistency = false

;[GPIO]
;IO0 = input disabled 10 false false 
;IO1 = input disabled 10 false false 
;IO3 = input disabled 10 false false 
;IO4 = built-in-led disabled 10 false false 
;IO12 = input-pullup disabled 10 false false 
;IO13 = input-pullup disabled 10 false false 
LEDType = WS2812
LEDNumbers = 2
LEDColor = 150 150 150

[AutoTimer]
Interval = 7

[DataLogging]
DataLogActive = false
DataFilesRetention = 3

[Debug]
LogLevel = 1
LogfilesRetention = 3

[System]
Tooltip = true
TimeZone = CET-1CEST,M3.5.0,M10.5.0/3
;TimeServer = pool.ntp.org
;Hostname = undefined
RSSIThreshold = -75
CPUFrequency = 160
SetupMode = false

[SybexX]

Have you tried to change CamGainceiling to e.g. x16?
That could help you too: #1949

[mt7479]

Thanks, I was not aware of the CamGainceling option. Set it to x16 and also got misreadings.
Cranked it to x32 and changed from greyscale to normal mode and this yields much better results.

However if the meter does not change, I turn it on temporarily because it's not the heating season yet, there are false readings but they correct themselves in subsequent readings (neg. rate).

Regrading dimming the Led: I actually have some polarizing foil at home which got me the best results on my other meter. I may case however it still required carefully positioning of the led. Do you think my last digits are to dark and that's the issue of the misreadings ?

[SybexX]

I rather assume that the border around the last digit is causing the problem.
Have you tried other models, like e.g. dig-class11_1900_s2_q.tflite oder dig-class11_1910_s2_q.tflite ?

[mt7479]

I did try the non-quantized versions dig-class11_1910_s2 and dig-class11_2000_s2 with the same results. However this was before your suggestion of increasing CamGainceling.

If I understand the docs correctly the non-quantized models are more accurate but take more processing power so I chose those instead of the q versions.

My idea would be to rerun with the other models (in normal picture mode), upping the CamGainceling in between if errors occur and moving to the next model if needed. Should I prefer the quantized versions over the non-quantized version for this attempt ?

[SybexX]

I can't actually make a precise recommendation for this, since the alignment and the calculation/processing take up most of the time and the evaluation of the digits is actually not significant and is also very dependent on the image quality.
The experience I have had is, if the image quality is very good(for the ESP32 and not for the human eye^^), the quantized model are faster if it is bad, a non-quantized model can be just as fast. So I always try both variants and take the best of the tested.
What the accuracy is also improved, look at the images that were used to train the model and then try to change the camera settings until you have almost the same image quality.