scene text competition

[wanghaisheng]

OCR的国际权威评测平台 ICDAR（International Conference on Document Analysis and Recognition）竞赛

在上一期中我们介绍了我们团队在OCR的国际权威评测平台ICDAR（International Conference on Document Analysis and Recognition）竞赛里所取得的佳绩，我们当时在ICDAR的互联网图片（Born-Digital Images）数据集上的两个任务（文本定位和单词识别）上都取得国际领先。最近，我们在ICDAR竞赛的另一个核心数据集：对焦自然场景图片（Focused Scene Text Images），也取得突破。以下详细介绍。

2.1 对焦自然场景图片里的文本定位任务比赛（Task1-Text Localization, ICDARFocused Scene Text Images） http://rrc.cvc.uab.es/?ch=2&com=evaluation&task=1

Focused Scene Text Image是用相机对准自然场景存在的文本拍摄得到的图像，这些文本包括海报、交通标志、告示牌、橱窗、店铺名称、衣服、铭牌等物体上的字符，文本定位任务就是确定图像中文本行的准确边界。该任务的训练集229幅，测试集233幅。由于自然场景中的文本定位和识别是OCR领域中的一个重要的研究方向，有一些研究机构和个人公布了自己收集和标注的数据集，通过搜集这些公开的数据集获得图像1560幅，作为补充训练集。在训练网络时，对训练集用了多种手段做了数据增强，实际训练集扩充到20000幅左右。我们的最新模型在该任务上取得了第一名的佳绩，如下图所示。

图3. 对焦自然场景图片里的文本定位任务比赛排名（http://rrc.cvc.uab.es/?ch=2&com=evaluation&task=1 ）

部分检测结果如下图所示，全部的检测结果可在网站上查询，网址：http://rrc.cvc.uab.es/?ch=2&com=evaluation&view=method_samples&task=1&m=30717&gtv=1

图4. 部分文本检测结果

2.2.对焦自然场景图片里的单词识别任务比赛（Task3-Word Recognition，ICDAR Focused Scene Text Images） http://rrc.cvc.uab.es/?ch=2&com=evaluation&task=2

Focused Scene Text Image单词识别任务需要在文本图像中抠出单词区域，四个边界向外扩展4个像素点，构成数据集，训练集848幅，测试集1095幅。在训练网络时，使用外部数据集约900万幅。采用CNN提取图像特征，采用RNN学习序列关系，并加入Attention机制以改善RNN的性能，进行识别。我们的最新模型在该任务上取得了第一名的佳绩，如下图所示。

图5. 对焦自然场景图片里的单词识别任务比赛排名（http://rrc.cvc.uab.es/?ch=2&com=evaluation&task=3 ）

部分单词如下图所示，这些单词在字体、尺寸、排列间距、倾斜、阴影、背景、模糊等方面都有变化，我们一方面增强网络结构以适应这些变化，另一方面有针对性的生成大量的合成样本用于训练网络，最终克服了这些不利因素，正确识别出单词。

图6. 部分单词图像

2.3.互联网场景图片里的文本定位任务比赛（Task1-Text Localization，ICDAR Born Digital Images）

近期，我们改进了用于互联网图片文本检测的网络结构，再一次刷新了互联网场景图片里的文本定位任务比赛上的记录，如下图所示。全部的检测结果可在网站上查询，网址：http://rrc.cvc.uab.es/?ch=1&com=evaluation&view=method_samples&task=1&m=30556>v=1

图7. 互联网图片文本检测任务上的排名（http://rrc.cvc.uab.es/?ch=1&com=evaluation&task=1 ）

2.4.互联网场景图片里的单词识别任务比赛（Task3-Word Recognition，ICDAR Born Digital Images）

我们改进了用于互联网图片单词识别的网络结构，加入Attention机制来改善RNN的性能，再一次刷新了互联网场景图片里的单词识别任务比赛上的记录，如下图所示。

[wanghaisheng]

http://u-pat.org/ICDAR2017/program_competitions.php ICDAR2017 Competitions

We are pleased to announce that the ICDAR2017 will organize a set of competitions dedicated to a large set of document analysis problems. You are cordially invited to participate to this scientific event that will be a very good opportunity to objectively compare the quality of algorithms on different categories of challenges. You will find below the different categories of competitions, and the URL of their respective website, that will allow you to get all the required information for participating: Category: Handwritten Historical Document Layout Recognition

cBAD: ICDAR2017 Competition on Baseline Detection
ICDAR2017 Competition on Layout Analysis for Challenging Medieval Manuscripts
ICDAR2017 Competition on Historical Book Analysis

Category: Historical Handwritten Script Analysis

ICDAR 2017 Competition on the Classification of Medieval Handwritings in Latin Script
ICDAR2017 Competition on Historical Document Writer Identification (Historical-WI)
Competition on Multi-script Writer Identification Using LAMIS-MSHD and CERUG Databases

Category: Character/Word Spotting

Competition on Query-by-Example Glyph Spotting of Southeast Asian Palm Leaf Manuscript Images
Handwritten Keyword Spotting Competition

Category: Handwriting Recognition

ICDAR2017 Competition on Handwritten Text Recognition on the READ Dataset
ICDAR2017 Competition on Information Extraction in Historical Handwritten Records

Category: Document Image Binarization

ICDAR2017 Competition on Document Image Binarization (DIBCO 2017)

Category: Document Recognition (Layout analysis & Text Recognition)

ICDAR2017 Competition on Recognition of Documents with Complex Layouts – RDCL2017
ICDAR2017 Competition on Recognition of Early Indian Printed Documents – REID2017
ICDAR2017 Competition on Page Object Detection

Category: Document Reconstruction

Smartphone-captured Document Image Reconstruction from Multiple Views

Category: Post OCR Correction

ICDAR2017 Competition on Post-OCR Text Correction

Category: Robust Reading Competitions

ICDAR2017 Competition on Reading Chinese Text in the Wild (RCTW-17)
ICDAR2017 Robust Reading Challenge on COCO-Text
ICDAR2017 Robust Reading Challenge on Text Extraction from Biomedical Literature Figures (DeTEXT)
ICDAR2017 Robust Reading Challenge on Omnidirectional Video
ICDAR2017 Robust Reading Challenge on Multi-lingual Scene Text Detection and Script Identification – RRC-MLT

Category: Text in Video

ICDAR2017 Competition on Arabic Text Detection and Recognition in Multi-resolution Video Frames
Competition on Video Script Identification

Category: Forensics

Competition on File Type Identification

Category: Miscellaneous Competitions

ICDAR2017 Competition on Multi-font and Multi-Size Digitally Represented Arabic Text

[wanghaisheng]

https://github.com/xylcbd/ocr-open-dataset/blob/8e73cb0e273d61607548cbf2cafafcc6377a4996/README.md

[wanghaisheng]

http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0126200

[wanghaisheng]

http://www.iapr-tc11.org/mediawiki/index.php/Datasets_List 别人整理的数据集 截止日期 2015年 Datasets List

From TC11

Jump to: navigation, search

NOTICE: TC11 datasets will be soon moved to the new Web portal at http://tc11.cvc.uab.es This page will remain available but will not be updated from January 2015 onwards.

Datasets -> Datasets List

Last updated: 2015-001-23

See the datasets sorted according to the Journal / Conference they first appeared in.

Contents

• 1 Complex Text Containers
  • 1.1 Scene Text
• 2 Machine-printed Documents
• 3 Graphical Documents
• 4 Mixed Content Documents
• 5 Handwritten Documents
  • 5.1 On-line and Off-line
  • 5.2 On-line
  • 5.3 Off-line
• 6 Software and Tools

Complex Text Containers

Scene Text

• MSRA Text Detection 500 Database (MSRA-TD500)
• The Street View Text Dataset
• The Street View House Numbers (SVHN) Dataset
• NEOCR: Natural Environment OCR Dataset
• KAIST Scene Text Database
• ICDAR 2003 Robust Reading Competitions
• ICDAR 2005 Robust Reading Competitions

Machine-printed Documents

• Table Ground Truth for the UW3 and UNLV datasets
• The DocLab Dataset for Evaluating Table Interpretation Methods
• [the IMPACT data base] The dataset contains more than half a million representative text-based images compiled by a number of major European libraries. Covering texts from as early as 1500, and containing material from newspapers, books, pamphlets and typewritten notes, the dataset is an invaluable resource for future research into imaging technology, OCR and language enrichment.
• PRImA Layout Analysis Dataset
• DFKI Dewarping Contest Dataset (CBDAR 2007) The dataset, that was used in the CBDAR 2007 Dewarping Contest, contains 102 camera captured documents with their corresponding ASCII text ground-truth. Additionally, text-line level ground-truth was also prepared to benchmark curled text-line segmentation algorithms. Part of the dataset (76 out of 102 pages) were also scanned with a flat-bed scanner to create a ground-truth image for image based evaluation of page dewarping algorithms.
• APTI: Arabic Printed Text Image Database
• LRDE Document Binarization Dataset (LRDE DBD) This dataset is composed of documents images extracted from the same French magazine: Le Nouvel Observateur, issue 2402, November 18th-24th, 2010. The provided dataset is composed of 375 Full-Document Images (A4 format, 300-dpi resolution).
• RETAS OCR Evaluation Dataset The RETAS dataset (used in the paper by Yalniz and Manmatha, ICDAR'11) is created to evaluate the optical character recognition (OCR) accuracy of real scanned books. The dataset contains real OCR outputs for 160 scanned books (100 English, 20 French, 20 German, 20 Spanish) downloaded from the Internet Archive website. The corresponding ground truth text for each scanned book is obtained from the Project Gutenberg database. The OCR output of each scanned book is aligned with its ground truth at the word and character level and the alignment output is provided along with estimated OCR accuracies. The dataset is provided for research purposes.

Graphical Documents

• Chem-Infty Dataset: A ground-truthed dataset of Chemical Structure Images
• Braille Dataset - Shiraz University
• TradeMarks Image Database - By way of Benoit Huet, 999 trademark and logo images

Mixed Content Documents

• Tobacco800 Document Image Database - composed of 1290 document images collected and scanned using a wide variety of equipment over time.

Handwritten Documents

On-line and Off-line

• ICDAR 2009 Signature Verification Competition (SigComp2009)

• ICFHR 2010 Signature Verification Competition (4NSigComp2010)

• ICDAR 2011 Signature Verification Competition (SigComp2011)

• ICFHR 2012 Signature Verification Competition (4NSigComp2012)

• CASIA Online and Offline Chinese Handwriting Databases - The Chinese handwriting datasets were produced by 1,020 writers using Anoto pen on papers, such that both online and offline data were obtained. Both the online and the offline dataset consists of three subsets for isolated characters (DB1.0–1.2, about 3.9 million samples of 7,356 classes) and three for handwritten texts (DB2.0–2.2, about 5,090 pages and 1.35 million characters). The datasets are free for academic research for handwritten document segmentation and retrieval, character and text line recognition, writer adaptation and identification.

• Persian Heritage Image Binarization Dataset (PHIBD 2012) This dataset contains 15 historical and old manuscript images collected from the historical records at the Documents and old manuscripts treasury of Mirza Mohammad Kazemaini (affiliated with Hazrate Emamzadeh Jafar), Yazd, Iran. The images suffer from various types of degradation including bleed-through, faded ink, and blur. The dataset is the first in a series to provide document images and their ground truth as a contribution to Document image analysis and recognition (DAIR) community. It is planned to provide more data and ground-truth information in the fture.

On-line

• CROHME: Competition on Recognition of Online Handwritten Mathematical Expressions

• Devanagari Character Dataset

• Harbin Institute of Technology Opening Recognition Corpus for Chinese Characters (HIT-OR3C)

• IAM Online Document Database (IAMonDo-database)

• IAM On-Line Handwriting Database

• UNIPEN database (Click on link 'CDROMs')

• Nakagawa Lab Online Handwriting Database

• The Informal Competition of Recognizing On-line Words (ICROW) by the Unipen Foundation

Off-line

• The Rimes Database comprises 12,723 handwritten pages corresponding to 5605 mails of two to three pages. It was collected by asking volunteers to write a letter given one of nine predefined scenarios related to business/customer relations. The dataset has been used in numerous competitions in ICDAR and ICFHR. It is available for research purposes only, through the Web site of the authors.

• IBN SINA: A database for research on processing and understanding of Arabic manuscripts images

• CEDAR Off-line Handwriting CDROM1

• CVL-Database - An Off-line Database for Writer Retrieval, Writer Identification and Word Spotting

• IAM Database - A full English sentence database for off-line handwriting recognition.

• The GERMANA Dataset - GERMANA is the result of digitising and annotating a 764-page Spanish manuscript entitled “Noticias y documentos relativos a Doña Germana de Foix, ́última Reina de Aragón", written in 1891 by Vicent Salvador. It contains approximately 21K text lines manually marked and transcribed by palaeography experts.

• The RODRIGO Dataset - RODRIGO is the result of digitising and annotating a manuscript dated 1545. Digitisation was done at 300dpi in color by the Spanish Culture Ministry. The original manuscript is a 853-page bound volume, entitled "Historia de España del arçobispo Don Rodrigo", completely written in old Castilian (Spanish) by a single author. Annotation exists for text blocks, lines and transcriptions, resulting in approximately 20K lines and 231K running words from a lexicon of 17K words.

• MARG- Medical Article Records Groundtruth - A freely-available repository of document page images and their associated textual and layout data. The data has been reviewed and corrected to establish its "ground truth". Please contact Dr. George Thoma (thoma@lhc.nlm.nih.gov) at the National Library of Medicine for more information.

• Hindi font samples by Andras Kornai, June 5 2003

Software and Tools

• Europeana API The Europeana network represents more than 2,500 cultural heritage organisations and is the principal point of reference for digitised European culture. Europeana offers open access to over 32 million records, a large percentage of which are document images originating from various memory institutions including national libraries and archives. In many cases it offers links to high resolution scans of such documents. The Europeana APIs allow you to search and retrieve the contents of their database for use in your own applications. Two APIs are offered. The first is a REST-API that is suited for dynamic search and retrieval of our data. The second API is more experimental and supports download of complete datasets and advanced semantic search and retrieval of our data via the SPARQL query language.
• Tools from the IMPACT Centre of Competence The tools offered by the Impact Centre of Competence are software components which have been developed by the different technical IMPACT partners during the IMPACT project (2008-2012). Generally, a "tool" is a piece of software which operates on image or text data modifying the data or extracting information from it. Every IMPACT tool has a specific functionality related to OCR, to the the pre- and post-processing stages. They include new approaches in areas such as image enhancement, segmentation, and document structuring, alongside existing and experimental OCR engines.
• GEDI: Groundtruthing Environment for Document Images - A generic annotation tool for scanned text documents.
• PixLabeler - a research tool for labeling elements in a document image at a pixel level.
• OCRopus(tm) - The OCRopus(tm) open source document analysis and OCR system
• The Hidden Markov Model Toolkit (HTK) - a portable toolkit for building and manipulating hidden Markov models
• Bidirectional Long-Short Term Memory Networks - Implementation of Bidirectional Long-Short Term Memory Networks (BLSTM) combined with Connectionist Temporal Classification (CTC) - including examples for Arabic recognition.
• SRILM - The SRI Language Modeling Toolkit - SRILM is a toolkit for building and applying statistical language models (LMs), primarily for use in speech recognition, statistical tagging and segmentation, and machine translation.
• Torch 5 - a Matlab-like environment for state-of-the-art machine learning algorithms.
• PRTools - a Matlab based toolbox for pattern recognition

[wanghaisheng]

https://github.com/xellows1305/Document-Image-Dewarping https://github.com/duerig/laser-dewarp https://github.com/xellows1305/Scene-Text-Rectification