search

PaddlePaddle / PaddleOCR

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)
https://paddlepaddle.github.io/PaddleOCR/
Apache License 2.0
43.79k stars 7.79k forks source link

What type of OCR is better for US plate recognition? #10413

Closed bit-scientist closed 5 months ago

bit-scientist commented 1 year ago

I am relatively new to the filed of OCR. I have been reading the book Dive into OCR provided here. It turns out the recogntion consists of two stages: text detection and text recognition. Well, there's also end-to-end approachs like PGNet, but they're not as good as 2-stage algorithms. I happen to work on US plate recogntion where plates vary so much in terms of font, color and shape.

See some samples below:

image

I should note that these plates are available on the Web and am not sure if they are used nowadays. The texts mostly appear horizontally, but some have vertical text as well. I have'nt seen any curved texts so far. What I am concerned about is what algorithms (text detector & text recognizer) work best on these types of plates. With all the available algorithms for OCR (both detection & recognition) here, I am quite puzzled which one to choose for my case. I hope you could direct me to a narrower path so that I don't spend too much time on what might not work at the end. Thanks.

ToddBear commented 1 year ago

Here is a model that has been trained for Chinese license plate recognition, and you can refer to his design for US license plate recognition model.

I suggest that you can fine-tune the PP-OCR v3 recognition model and PP-OCR v3 detection model on the US license plate data.

https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/applications/%E8%BD%BB%E9%87%8F%E7%BA%A7%E8%BD%A6%E7%89%8C%E8%AF%86%E5%88%AB.md

bit-scientist commented 1 year ago

Thanks a ton @ToddBear. It looks like PP-OCRv3 worked quite well on CCPD. The part I'm struggling to understand: is PP-OCRv3 itself detection (and recogntion) model? Is PP-OCRv3 detailed in this paper? Thanks again.

UserWangZz commented 5 months ago

This issue has not been updated for a long time. This issue is temporarily closed and can be reopened if necessary.