Dark Background - works in demo

[Gianfilippo980]

Te image that I need to OCR looks like this (it is a time indication):

Of course, I tested it in the demo and it works flawlessly:

Then I installed the software on my Raspberry Pi and I can't get it to replicate the results:

Python 3.11.2 (main, Nov 30 2024, 21:22:50) [GCC 12.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import easyocr
>>> reader = easyocr.Reader(['en'])
Neither CUDA nor MPS are available - defaulting to CPU. Note: This module is much faster with a GPU.
>>> reader.readtext('zona_orario.png')
[([[np.int32(21), np.int32(3)], [np.int32(43), np.int32(3)], [np.int32(43), np.int32(21)], [np.int32(21), np.int32(21)]], '0', np.float64(0.9998962906071398)), ([[np.int32(55), np.int32(5)], [np.int32(89), np.int32(5)], [np.int32(89), np.int32(21)], [np.int32(55), np.int32(21)]], 'GED', np.float64(0.0))]
>>> reader.readtext('zona_orario.png', allowlist='0123456789')
[([[np.int32(21), np.int32(3)], [np.int32(43), np.int32(3)], [np.int32(43), np.int32(21)], [np.int32(21), np.int32(21)]], '0', np.float64(0.9999812842291114)), ([[np.int32(55), np.int32(5)], [np.int32(89), np.int32(5)], [np.int32(89), np.int32(21)], [np.int32(55), np.int32(21)]], '77', np.float64(0.012275525823156254))]
>>> reader.readtext('zona_orario.png', allowlist='0123456789', mag_ratio=3)
[([[np.int32(54), np.int32(3)], [np.int32(88), np.int32(3)], [np.int32(88), np.int32(20)], [np.int32(54), np.int32(20)]], '13', np.float64(0.999999915706304))]
>>> reader.readtext('zona_orario.png', allowlist='0123456789', mag_ratio=2)
[([[np.int32(21), np.int32(5)], [np.int32(42), np.int32(5)], [np.int32(42), np.int32(20)], [np.int32(21), np.int32(20)]], '0', np.float64(0.9999947547981378)), ([[np.int32(54), np.int32(5)], [np.int32(89), np.int32(5)], [np.int32(89), np.int32(20)], [np.int32(54), np.int32(20)]], '2', np.float64(0.03254324029984801))]
>>> reader.readtext('zona_orario.png', allowlist='0123456789', mag_ratio=2.5)
[([[np.int32(55), np.int32(4)], [np.int32(89), np.int32(4)], [np.int32(89), np.int32(19)], [np.int32(55), np.int32(19)]], '77', np.float64(0.011731833376220437))]
>>> reader.recognize('zona_orario.png', allowlist='0123456789')
[([[0, 0], [94, 0], [94, 25], [0, 25]], '77', np.float64(0.01394686076874571))]
>>> 

As you can see, if I set a higher magnification the minutes are read correctly, but then the hours are not. Any advice?

[Gianfilippo980]

By the way, I c drew the boxes that were parsed in the first attempt, and those look pretty good, just like the ones in the demo:

[Gianfilippo980]

Inverting the colors doesn't seem to help:

>>> reader.recognize('orario_invertito.png', allowlist='0123456789.')
[([[0, 0], [94, 0], [94, 25], [0, 25]], '77', np.float64(0.013375049105307402))]
>>> reader.readtext('orario_invertito.png', allowlist='0123456789.')
[([[np.int32(55), np.int32(5)], [np.int32(89), np.int32(5)], [np.int32(89), np.int32(21)], [np.int32(55), np.int32(21)]], '77', np.float64(0.010336355524602742))]