CheckDigitIncreaseConsistency not working for model dig-cont_0712_s3_q.tflite

[fsck-block]

The Problem

I'm using a meter with a very early analog to digital transition
This results (despite other problems) to inaccuracy when the analog pointer reaches values from .8 to .99 and the lowest digit is nearly switched to the next one.

Note: Meter shows value 232.91

The parameter CheckDigitIncreaseConsistency promises some help,

An additional consistency check. It especially improves the zero crossing check between digits.

but the check is skipped caused by wrong model

checkDigitIncreaseConsistency = true - ignored due to wrong CNN-Type (not Digit Classification)

Using a other digit models (e.g dig-class11_1701_s2.tflite) is not an option, because they have a worse recoginition rate with my meter.

Is there a reason why the continuous output networks models are ignored?

Version

15.7.0

Logfile

n/a

Expected Behavior

Parameter CheckDigitIncreaseConsistency is effective for all digital models

Screenshots

No response

Additional Context

No response

[SybexX]

As far as i know, this option is only being available for digital ROIs, but you also activated analog ROIs.

[SybexX]

https://jomjol.github.io/AI-on-the-edge-device-docs/Correction%20Algorithm/#checkdigitincreaseconsistency

[fsck-block]

I'm talking exactly about the that, what the description of the correction algorithm describes.
This is exactly what I need

But unfortunately this not available for the digital ROI model "dig-cont_0712_s3_q.tflite"

Have a look at jomjol_flowcontroll/ClassFlowPostProcessing.cpp@869

if (NUMBERS[j]->checkDigitIncreaseConsistency) {
    if (flowDigit) {
        if (flowDigit->getCNNType() != Digit) {
            ESP_LOGD(TAG, "checkDigitIncreaseConsistency = true - ignored due to wrong CNN-Type (not Digit Classification)");
        }

The model "dig-cont_0712_s3_q.tflite" has CNNType DoubleHyprid10 not Digit even it classifies "Digits", as far as I understand.

[SybexX]

With DoubleHyprid10, a different/expanded calculation is carried out, but depending on AnalogDigitTransitionStart.
https://jomjol.github.io/AI-on-the-edge-device-docs/Watermeter-specific-analog---digit-transition/
https://jomjol.github.io/AI-on-the-edge-device-docs/Correction%20Algorithm/#analogdigittransitionstart

[fsck-block]

Yes, setting AnalogDigitTransitionStart to 6.0 helps a little bit, but even with this setting problems occur on zero-crossing.

Made a development version of the software with activated CheckDigitIncreaseConsistency for DoubleHyprid10 model.
Zero-crossing works well, CheckDigitIncreaseConsistency corrects the raw value (within range 232.98 to 232.99).

As far as i understand the CheckDigitIncreaseConsistency is a subsequent calculation, so it might be usefull for all models classifying digits.

[SybexX]

Simply put the last two digits up a bit, then they should be recognized as 2.8 or 2.9.

[fsck-block]

Simply put the last two digits up a bit, then they should be recognized as 2.8 or 2.9.

Great, you are right. Moving the digit ROI up by four pixel leads to a perfect classification during zero-crossing.

So my original problem is solved, I no longer need the calculation of CheckDigitIncreaseConsistency.

Thanks