OCR

Optical Character Recognition

There are two parts in OCR.

1. Text Detection
2. Text Recognition

Text Detection

Paper related

Scene Text Detection with Polygon Offsetting and Border Augmentation

Regression-based

SSD: Single shot detector https://arxiv.org/abs/1512.02325

TextBoxes https://arxiv.org/abs/1611.06779

Segmentation-based

in today's paper, it is popular and can be trained stably.
PixelLink

End to End

including Text Detection and Text Recognition
FOTS

vs Object Detection

OCR is more complex than object detection

the reason is
1. high density
2. similar characters; i, I, T, l
3. a lot of languages

Text Recognition

CRNN
GRCNN

code

keras-ocr

https://keras-ocr.readthedocs.io/en/latest/index.html

하지만 밑의 github을 추천한다.
https://github.com/faustomorales/keras-ocr

tesseract

참고 자료

https://tv.naver.com/v/4578167
https://github.com/clovaai/CRAFT-pytorch
https://blogs.sas.com/content/saskorea/2018/12/21/딥러닝을-활용한-객체-탐지-알고리즘-이해하기/