parksunwoo / memo-archive

memo of dev issues
0 stars 0 forks source link

Scene Text Detection and Recognition: Recent Advances and Future Trends #1

Closed parksunwoo closed 3 years ago

parksunwoo commented 5 years ago

OCR 기본개념 및 연구동향, 데이터셋 및 알고리즘별 장단점 비교

1.Introduction

2.Recent Advances in Scene Text Detection and Recognition

2.1 text detection

texture based method

텍스트의 특성을 사용해서, local intensities, filter responses and wavelet coefficients, to distinguish between text and non text areas in the images

component based method

주로 사용되는 방법으로 다양한 방법을 통해 후보구성요소를 추출하고 필터링을 통해 비-문자형 요소를 제거 (수동규칙 또는 훈련된 classifier)

hybrid methods

texture based & component based

Table 1. Comparison of existing text detection methods. Strength ::

Weakness ::

2.2 text recognition

문서이미지와는 달리 일반 이미지에서 문자인식하는 부분에 몇가지 장애물이 있는데. 잘못된 다수의 알람, gibberish bottom-up and top-down information integrated in a unified model conditional random field, tolerate errors in character detection

part based text recognition

robust to nose, blur, partial occlusion and font variation but depends on the detailed annotation information. including character models and part annotations

Table 2 Comparison of existing text recognition methods.

strength::

weakness::

2.3 end to end text recognition

일반적인 object detection, 단어는 특별한 종류의 물체, character는 단어의 일부분 주어진 단어 리스트에서만 가능, 단어리스트가 주어지지 않은 이미지에는 적용불가

strength::

weakness::

2.4 APP, products

3. datasets

ICDAR OSTD MSRA-TD500 SVT NEOCR KAIST Chars74K SVHN IIIT 5K-Word

3.2 Evaluation Protocols

3.2.1 Evaluation Protocols for Text Detection Algorithms

4. discussion

5. conclusions

Therefore, detecting and recognizing texts in natural scenes have become important and vibrant research areas in computer vision