Very poor recognition despite good image

[TheNavy]

Does anyone have an idea how I can improve the picture from my ESP-CAM. Currently I have to reset the "preValue" every 1-2 days because it sometimes detects the wrong numbers for several hours and then the rate is too high. If I set the "MaxRateValue" higher he often detects too high values.
I already have external LEDs in use and feel the image is actually good (except for the last digit but I do not care).

Here is an Screenshot from the UI:
https://i.ibb.co/Yh1LS4X/aiontheedge.png

Thanks :)

[jomjol]

Cannot load the image. Can you update the link?

[Impact123]

The URL itself still worksworked, just copy & paste it as text into the address bar. For your convenience:

• OP's link: https://ibb.co/pdMLT6Q
• Direct link to image: https://i.ibb.co/Yh1LS4X/aiontheedge.png

As Attachment

[TheNavy]

Sorry my bad,
thanks for the correction of the link :)

Edit:
I've played with the LED Brightness a little bit but still:

[jomjol]

The image quality is really good (at least until the first subdigit. You can optimize your ROIs a bit: https://github.com/jomjol/AI-on-the-edge-device/wiki/ROI-Configuration

[TheNavy]

I have now changed the ROI's several times (pixel by pixel) but still have problems with it every 1-2 days.
Interestingly, the problem occurs mostly on the 5th ROI.
Here is a screenshot from the wizard:

[jomjol]

Then the only solution is to integrate your images into the training (see: https://github.com/jomjol/AI-on-the-edge-device/wiki/ROI-Configuration)

[wernight]

I don't have a great visual quality and it's already hard to get this quality due to the very limited space. However I wish to help find ways to improve the recognition as a human could very clearly identify the numbers.

Also having a similar issue with one where I could get a better visual quality:

For the second one I got better results by limiting the area for each digit to avoid the black bottom/top. However this doesn't follow the wiki suggestions.

[wernight]

Some ideas:

• Allow to manually set a brightness/contrast (I think my hardware doesn't support it). Or even set some convolution matrix.
• Allow to hide/block out part of the image.
• Make it easier to manually train a network.