Recognition of non-English characters

[psloplion]

Hello. I have created my own dataset for Ukrainian character recognition. Added symbols to the dictionary. For training, I used the CRNN model. But during training, the recall and precision always show 0.0000. When testing the model, predictions are not shown, but instead of them, puffy values. What could it be? Thanks.

[gaotongxiao]

Did you add Ukrainian characters to the dictionary as well?

And for the evaluation - now the evaluation metric only takes Chinese and English characters into account. Don't forget to add Ukranian character set to valid_symbols for the correct evaluation result: https://mmocr.readthedocs.io/en/dev-1.x/api/generated/mmocr.evaluation.metrics.WordMetric.html#wordmetric

[psloplion]

Do I understand correctly that you can add the valid_symbols line to the dict in the base /default_runtime.py file here?

[psloplion]

Yes, I added all Ukrainian characters to the new file. Assembly files look like this:

[gaotongxiao]

Do I understand correctly that you can add the valid_symbols line to the dict in the base /default_runtime.py file here?

Yes, and you should add the range of Ukrainian character's Unicode to valid_symbol in a way similar to Chinese characters (FYI, Chinese characters range from \u4e00t to \u9fa5):

valid_symbol: str = '[^A-Z^a-z^0-9^\u4e00-\u9fa5]',

The implementation can be found here: https://mmocr.readthedocs.io/en/dev-1.x/_modules/mmocr/evaluation/metrics/recog_metric.html#WordMetric