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/
6.03k stars 651 forks source link

Low-contrast digits recongition #603

Closed pcmihnea closed 2 years ago

pcmihnea commented 2 years ago

Hi,

Due to physical contraints and water meter design, I am unable to obtain a higher contrast image of a watermeter with digits only display.

image The meter in question has red-ish digits on a white background ( https://metra-su.cz/en/product/dry-dial-vane-wheel-single-jet-water-meter-js-02-smart ). There's no easy way to mount the camera directly due to it's position inside a well and the attached wireless module ( https://metra-su.cz/en/product/electronic-radio-module-e-rm-30-for-residential-water-meters-apator-powogaz ).

The highest readability I could obtain was with a 8x WS2812 LEDs ring ( https://uk.pi-supply.com/products/8-led-neopixels-ring ) running full green only. The ESP32-CAM has the default OV2640 camera module. I've used yellow labels (with '-30%' markings) for maximum contrast with the dark background.

How may I improve recongition of the meter's readings? image image image image image

jomjol commented 2 years ago

In this case it is really difficult. You most propably would need a dedicated trained network. If you are familiar with Python and Tensorflow you can find details here: https://github.com/jomjol/neural-network-digital-counter-readout

But to be hones, I strongly recommend to get closer to the meter. I also have a system inside a metal box and used an external WLAN antenna. That works very well.

pcmihnea commented 2 years ago

Tried different placement and lighting - moved the camera as close as possible, replaced the LED ring with the onboard one, and trimed the lens focus. Still no luck... EDIT: the camera brightness is configured to -2, to prevent over-saturation. image

jomjol commented 2 years ago

Well it might be, that it is needed to implement your numbers in one of the next trainings rounds - for details see here: https://github.com/jomjol/AI-on-the-edge-device/wiki/ROI-Configuration