Adding New Characeters To Text Recog models

[TheSeriousProgrammer]

I am currenly using CRNN's text recog model, but it doesnt have a few special charecters which I need such as ("?","/") and a few others
How can I calibrate CRNN and modify my inference code to support these new charecters?

Thank you so much for the impressive framework though

[gaotongxiao]

Similar to #1488, it depends on the version that you're using.

MMOCR 0.x

You can control the supported character set by the parameter dict_type here:

mmocr/configs/_base_/recog_models/crnn.py

Lines 1 to 2 in 3188e53

	label_convertor = dict(
	type='CTCConvertor', dict_type='DICT36', with_unknown=False, lower=True)

We have a list of pre-defined DICTs, where DICT90 shall satisfy your requirement:

mmocr/mmocr/models/textrecog/convertors/base.py

Lines 24 to 26 in 37833ad

	DICT90=tuple('0123456789abcdefghijklmnopqrstuvwxyz'
	'ABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&\'()'
	'*+,-./:;<=>?@[\\]_`~'),

MMOCR 1.0

It can be controlled via the dict_file parameter:

mmocr/configs/textrecog/crnn/_base_crnn_mini-vgg.py

Lines 1 to 4 in f30c16c

	dictionary = dict(
	type='Dictionary',
	dict_file='{{ fileDirname }}/../../../dicts/lower_english_digits.txt',
	with_padding=True)

MMOCR 1.0 has provided some commonly used dictionary files, including english_digits_symbols that works for your case.

[TheSeriousProgrammer]

Thank you that's what I was looking for , will check it out and comeback

[TheSeriousProgrammer]

It worked perfectly , thanks for the response