Scene Text Recognition Recommendations

Everything about Scene Text Recognition

SOTA • Papers • Datasets • Code • Our Framework

Contents

• 1. Papers
• 2. Datasets
  • 2.1 Synthetic Datasets
  • 2.2 Benchmarks
• 3. Public Code
  • 3.1 Frameworks
  • 3.2 Algorithms
• 4. SOTA

---

1. Papers

All Papers Can be Find Here

• Latest Papers:

up to (2022-9-20)

• ECCV-2022:Levenshtein OCR
• ECCV-2022:Multi-Granularity Prediction for Scene Text Recognition
• arXiv-2022:A Scene-Text Synthesis Engine Achieved Through Learning from Decomposed Real-World Data
• arXiv-2022:Scene Text Recognition with Single-Point Decoding Network
• ECCV-2022-Technical-Report:Vision-Language Adaptive Mutual Decoder for OOV-STR
• WACV-2023:Seq-UPS: Sequential Uncertainty-aware Pseudo-label Selection for Semi-Supervised Text Recognition
• ECCV-2022-Technical-Report:1st Place Solution to ECCV 2022 Challenge on Out of Vocabulary Scene Text Understanding: End-to-End Recognition of Out of Vocabulary Words
• ECCV-2022-Technical-Report:Runner-Up Solution to ECCV 2022 Challenge on Out of Vocabulary Scene Text Understanding: Cropped Word Recognition

up to (2022-8-9)

• ECCV-2022:Toward Understanding WordArt: Corner-Guided Transformer for Scene Text Recognition

up to (2022-7-24)

• ECCV-2022:SGBANet: Semantic GAN and Balanced Attention Network for Arbitrarily Oriented Scene Text Recognition
• ECCV-2022:Scene Text Recognition with Permuted Autoregressive Sequence Models

up to (2022-7-9)

• arXiv-2022:MaskOCR: Text Recognition with Masked Encoder-Decoder Pretraining
• ACM-MM22:Reading and Writing: Discriminative and Generative Modeling for Self-Supervised Text Recognition

up to (2022-5-12)

• arXiv-2022:Multimodal Semi-Supervised Learning for Text Recognition
• IJCAI-2022:SVTR: Scene Text Recognition with a Single Visual Model

up to (2022-3-17)

• arXiv-2022:Training Protocol Matters: Towards Accurate Scene Text Recognition via Training Protocol Searching
  • 通过搜索训练参数来提升现有模型性能

up to (2022-3-11)

• arXiv-2022:Towards Open-Set Text Recognition via Label-to-Prototype Learning
  • 当测试阶段遇到训练集中没有出现过的字符时，应该如何应对，场景文字识别中的开集问题
• arXiv-2022:Text-DIAE: Degradation Invariant Autoencoders for Text Recognition and Document Enhancement
• arXiv-2022:Invariant Autoencoders for Text Recognition and Document Enhancement]

up to (2022-3-10)

• arXiv-2020:Hamming OCR: A Locality Sensitive Hashing Neural Network for Scene Text Recognition
  • 当识别种类数增大时，softmax embedding层就会更大，计算量也就会增大。本文提出使用汉明编码来进行解码，而不是使用one-hot进行解码

up to (2022-3-2)

• AAAI-2022:FedOCR: Efficient and Secure Federated Learning for Scene Text Recognition
  • 联邦学习用于场景文字识别
• AAAI-2022:Context-based Contrastive Learning for Scene Text Recognition
  • 对比学习用于场景文字识别

up to (2022-2-26)

• BMCV-2021:An Adaptive Rectification Model for Arbitrary-Shaped Scene Text Recognition
  • 提出新的矫正方法，在弯曲文本上效果好于TPS和MORAN

2. Datasets

All Datasets Can be Find Here

2.1 Synthetic Training Datasets

Dataset	Description	Examples	BaiduNetdisk link
SynthText	9 million synthetic text instance images from a set of 90k common English words. Words are rendered onto nartural images with random transformations		Scene text datasets(提取码:emco)
MJSynth	6 million synthetic text instances. It's a generation of SynthText.		Scene text datasets(提取码:emco)

---

2.2 Benchmarks

Dataset	Description	Examples	BaiduNetdisk link
IIIT5k-Words(IIIT5K)	3000 test images instances. Take from street scenes and from originally-digital images		Scene text datasets(提取码:emco)
Street View Text(SVT)	647 test images instances. Some images are severely corrupted by noise, blur, and low resolution		Scene text datasets(提取码:emco)
StreetViewText-Perspective(SVT-P)	639 test images instances. It is specifically designed to evaluate perspective distorted textrecognition. It is built based on the original SVT dataset by selecting the images at the sameaddress on Google Street View but with different view angles. Therefore, most text instancesare heavily distorted by the non-frontal view angle.		Scene text datasets(提取码:emco)
ICDAR 2003(IC03)	867 test image instances		Scene text datasets(提取码:mfir)
ICDAR 2013(IC13)	1015 test images instances		Scene text datasets(提取码:emco)
ICDAR 2015(IC15)	2077 test images instances. As text images were taken by Google Glasses without ensuringthe image quality, most of the text is very small, blurred, and multi-oriented		Scene text datasets(提取码:emco)
CUTE80(CUTE)	288 It focuses on curved text recognition. Most images in CUTE have acomplex background, perspective distortion, and poor resolution		Scene text datasets(提取码:emco)

---

2.3 Other Real Datasets

• The Real Datasets refer to this repo ku21fan/STR-Fewer-Labels

Dataset	Description	Examples	BaiduNetdisk link
COCO-Text	39K Created from the MS COCO dataset. As the MS COCO dataset is not intended to capture text. COCO contains many occluded or low-resolution texts		Others(提取码:DLVC)
RCTW	8186 in English. RCTW is created for Reading Chinese Text in the Wild competition. We select those in english		Others(提取码:DLVC)
Uber-Text	92K. Collecetd from Bing Maps Streetside. Many are house number, and some are text on signboards		Others(提取码:DLVC)
Art	29K. Art is created to recognize Arbitrary-shaped Text. Many are perspective or curved texts. It also includes Totaltext and CTW1500, which contain many rotated or curved texts		Others(提取码:DLVC)
LSVT	34K in English. LSVT is a Large-scale Streeet View Text dataset, collected from streets in China. We select those in english		Others(提取码:DLVC)
MLT19	46K in English. MLT19 is created to recognize Multi-Lingual Text. It consists of seven languages:Arabic, Latin, Chinese, Japanese, Korean, Bangla, and Hindi. We select those in english		Others(提取码:DLVC)
ReCTS	23K in English. ReCTS is created for the Reading Chinese Text on Signboard competition. It contains many irregular texts arranged in various layouts or written with unique fonts. We select those in english		Others(提取码:DLVC)

3 Public Code

3.1 Frameworks

PaddleOCR (百度)

• PaddlePaddle/PaddleOCR
• 特性 (截取至PaddleOCR)：
  • 使用百度自研深度学习框架PaddlePaddle搭建
  • PP-OCR系列高质量预训练模型，准确的识别效果
    • 超轻量PP-OCRv2系列：检测（3.1M）+ 方向分类器（1.4M）+ 识别（8.5M）= 13.0M
    • 超轻量PP-OCR mobile移动端系列：检测（3.0M）+方向分类器（1.4M）+ 识别（5.0M）= 9.4M
    • 通用PPOCR server系列：检测（47.1M）+方向分类器（1.4M）+ 识别（94.9M）= 143.4M
    • 支持中英文数字组合识别、竖排文本识别、长文本识别
    • 支持多语言识别：韩语、日语、德语、法语
    • 丰富易用的OCR相关工具组件
  • 半自动数据标注工具PPOCRLabel：支持快速高效的数据标注
    • 数据合成工具Style-Text：批量合成大量与目标场景类似的图像
    • 文档分析能力PP-Structure：版面分析与表格识别
    • 支持用户自定义训练，提供丰富的预测推理部署方案
    • 支持PIP快速安装使用
    • 可运行于Linux、Windows、MacOS等多种系统
• 支持算法(识别):
  • CRNN
  • Rosetta
  • STAR-Net
  • RARE
  • SRN
  • NRTR

---

MMOCR (商汤)

• open-mmlab/mmocr
• 特性(截取至MMOCR):
  • MMOCR 是基于 PyTorch 和 mmdetection 的开源工具箱，专注于文本检测，文本识别以及相应的下游任务，如关键信息提取。 它是 OpenMMLab 项目的一部分。
  • 该工具箱不仅支持文本检测和文本识别，还支持其下游任务，例如关键信息提取。
• 支持算法(识别)
  • CRNN (TPAMI'2016)
  • NRTR (ICDAR'2019)
  • RobustScanner (ECCV'2020)
  • SAR (AAAI'2019)
  • SATRN (CVPR'2020 Workshop on Text and Documents in the Deep Learning Era)
  • SegOCR (Manuscript'2021)

---

Deep Text Recognition Benchmark (ClovaAI)

• clovaai/deep-text-recognition-benchmark
• 特性:
  • Offical Pytorch implementation of What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis
  • 可自定义四阶段组件，如CRNN，ASTER
  • 容易上手，推荐使用

---

DAVAR-Lab-OCR (海康威视)

• hikopensource/DAVAR-Lab-OCR
• 特性:
  • 基于mmocr搭建，复现了一些算法，同时将来会用于海康自研算法开源
• 支持算法(识别)
  • Attention(CVPR 2016)
  • CRNN(TPAMI 2017)
  • ACE(CVPR 2019)
  • SPIN(AAAI 2021)
  • RF-Learning(ICDAR 2021)

---

3.2. Algorithms

CRNN

• Lua, Offical, 1.9k⭐: bgshih/crnn
  • 官方实现版本，使用Lua
• Pytorch, 1.9k⭐: meijeru/crnn.pytorch
  • 推荐使用🀄
• Tensorflow, 972⭐:MaybeShewill-CV/CRNN_Tensorflow
• Pytorch, 1.4k⭐:Sierkinhance/CRNN_Chinese_Characters_Rec
  • 用于中文识别版本的CRNN

---

ASTER

• Tensorflow, official, 651⭐: bgshih/aster
  • 官方实现版本，使用Tensorflow
• Pytorch, 535⭐:ayumuymk/aster.pytorch
  • Pytorch版本，准确率相较原文有明显提升

---

MORANv2

• Pytorch, official, 572⭐:Canjie-Luo/MORAN_v2
  • MORAN v2版本。更加稳定的单阶段训练，更换ResNet做backbone，使用双向解码器

---

4. SOTAs

All the models are evaluated in a lexicon-free manner

		Regular Dataset				Irregular  dataset
Model	Year	IIIT	SVT	IC13(857)	IC13(1015)	IC15(1811)	IC15(2077)	SVTP	CUTE
CRNN	2015	78.2	80.8	-	86.7	-	-	-	-
ASTER(L2R)	2015	92.67	91.16	-	90.74	76.1	-	78.76	76.39
CombBest	2019	87.9	87.5	93.6	92.3	77.6	71.8	79.2	74
ESIR	2019	93.3	90.2	-	91.3	-	76.9	79.6	83.3
SE-ASTER	2020	93.8	89.6	-	92.8	80		81.4	83.6
DAN	2020	94.3	89.2	-	93.9	-	74.5	80	84.4
RobustScanner	2020	95.3	88.1	-	94.8	-	77.1	79.5	90.3
AutoSTR	2020	94.7	90.9	-	94.2	81.8	-	81.7	-
Yang et al.	2020	94.7	88.9	-	93.2	79.5	77.1	80.9	85.4
SATRN	2020	92.8	91.3	-	94.1	-	79	86.5	87.8
SRN	2020	94.8	91.5	95.5	-	82.7	-	85.1	87.8
GA-SPIN	2021	95.2	90.9	-	94.8	82.8	79.5	83.2	87.5
PREN2D	2021	95.6	94	96.4	-	83	-	87.6	91.7
Bhunia et al.	2021	95.2	92.2	-	95.5	-	84	85.7	89.7
Luo et al.	2021	95.6	90.6	-	96.0	83.9	81.4	85.1	91.3
VisionLAN	2021	95.8	91.7	95.7	-	83.7	-	86	88.5
ABINet	2021	96.2	93.5	97.4	-	86.0	-	89.3	89.2
MATRN	2021	96.7	94.9	97.9	95.8	86.6	82.9	90.5	94.1

Baek's Reimplementation Version