Configuring ALPR (and paddleOCR) for Korean characters.

[techgeniey]

Hello, I am using codeproject.ai + AgentDVR combo to detect vehicles as well as license plates.
I am having fine with these except that the license plates are coming out but not recognizing korean characters.

Korean plates are usually in forms of nnn X nnnn or nn X nnnn. Problem is that numerals are recognize relatively well but
it's putting some numerals in place of Korean characters.

I have fiddled with installed paddleOCR even within the venv and when tested against korean plates, paddleocr itself recognize koreans very well and puts out correct results.

I am running this on Mac Mini M1 (so MacOS), and so far impressed with the results except this Korean character. Am I missing something here? I have changed modulesetting.json and added "COUNTRY" : "kr" under environment variable and changed following in options.py

    self.language              = 'korean'
    self.algorithm             = 'CRNN'
    self.cls_model_dir         = 'paddleocr/ch_ppocr_mobile_v2.0_cls_infer'
    self.det_model_dir         = 'paddleocr/Multilingual_PP-OCRv3_det_infer'
    self.rec_model_dir         = 'paddleocr/korean_PP-OCRv3_rec_infer'
    self.rec_char_dict_path    = 'bin/macos/python38/venv/lib/python3.8/site-packages/paddleocr/ppocr/utils/dict/korean_dict.txt'

I made sure of downloading both multi lingual detection and korean recognition inference modes and they are in place and ALPR module isn't complaining when started up. Still, I don't get Korean back. Does the module have any filtering sort of things in place? What's the correct github repository for this project?

Thanks so much for providing this wonderful package if Mike is seeing this post. But I just have this itch that I cannot seem to scratch myself..... :(

[MikeLud]

There is a filter
Try change line 245 to this
pattern = re.compile('[^a-zA-Z0-9\uAC00-\uD7A3\u1100-\u11FF\uA960-\uA97F\uD7B0-\uD7FF]+')

[techgeniey]

Thank you! I just figured out in my own but I really appreciate it! Wow! This is good!

[MikeLud]

What changes did you make to make it work?

[techgeniey]

I am sorry about the late reply. I simply changed the pattern similar to what you suggested.
pattern = re.comple('[^a-zA-Z0-9가-힣]+').
I don't know if you can see Korean characters but it's the first syllabel of Korean character set through the last possible syllabel
of Korean. I think your suggestion might be more extensive but Korean plates only use "syllabel" i.e. complete character combo (like consonant + vowel [ +consonant]) set.
Unless Korean authority change the rule to include individual character component, it's safe to use that component for now.

Unfornuately, ALPR still comes up short about 50% of the time. It sometimes incorrectly recognize Korean characters. Other times it has this weird habit of fragmenting license number into two and reverses the set. I think I might have to work on fine tuning or simply use the portion of the number to match the license number.

Thanks again for helping me and following up!