chullhwan-song / ocr

OCR research

OCR

Recognition - SOTA Performance (only unconstrained lexicon-free)

Evaluation

• TTM : Total-Text (multi-oriented)
• TTC : Total-Text (curved)
• 학습셋
  • SK : Synth90K
  • ST : SynthText

Paper	year	SVT	IIIT5k	IC03	IC13	SVTP	CUTE80	IC15	TTM	TTC	비고	학습셋
SOTA		91.5	94.0	96.7	95.8	86.6	88.5	79.4	76.3	66.7
[1] CRNN	15.07	80.8	78.2	89.4	86.7	66.8	54.9				base:CRNN	SK
[10] RARE	16.x	81.9	81.9	90.1	88.6	71.8					Rectification	SK
[9] STAR-Net	16.x	83.6	83.3	89.9	89.1	73.5					Rectification	SK
MY	17.x	79.91	84.26	89.5	90.82						Rectification	SK
[3]	18.x	87.1	89.4	94.7	94.0	73.9	62.5				GAN
[4] ASTER	18.x	89.5	93.4	94.5	91.8	78.5	79.5	76.1			Rectification
[5] AON	18.03	82.8	87.0	91.5		73.0	76.8	68.2
[6]	18.05	87.5	88.3	94.6	94.4			73.9
[2]	18.12	88.6	94.0	93.6	93.2	80.6	88.5	77.1	76.3	66.7		SK+ST
[7] ESIR	18.12	90.2	93.3		91.3	79.6	83.3	76.9			Rectification	SK+ST
[8] MORAN V1	19.01	88.3	91.2	95.0	92.4	76.1	77.4				Rectification	SK+ST
[8] MORAN V2	19.01	88.3	93.4	94.2	93.2	79.7	81.9				Rectification	SK+ST
[11] NRTR	19.10	91.5	90.1	95.4	95.8	86.6	80.9	79.4			Transformer
[12] SATRN	19.10	91.3	92.8	96.7	94.1	86.5	87.8	79.0			Transformer

• MY : MY implementation(2017), CRNN(modified networks)+jointly(CTC+Attention)+Rectification
• [1] An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
• [2] Recurrent Calibration Network for Irregular Text Recognition
• [3] Synthetically Supervised Feature Learning for Scene Text Recognition : [paper][review]
• [4] ASTER: An Attentional Scene Text Recognizer with Flexible Rectification # Rectification
• [5] Arbitrarily-oriented text recognition
• [6] Edit Probability for Scene Text Recognition
• [7] ESIR: End-to-end Scene Text Recognition via Iterative Rectification # Rectification
• [8] MORAN: A Multi-Object Rectified Attention Network for Scene Text Recognition # Rectification
• [9] STAR-Net: A SpaTial Attention Residue Network for Scene Text Recognition
• [10] RARE: Robust Scene Text Recognition with Automatic Rectification
• [11] NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition # self attention
• [12] SATRN: On Recognizing Texts of Arbitrary Shapes with 2D Self-attention # self attention

Detection

Papers's training Information : 링크

Evaluation

• R : Recall
• P : Precision
• F : F-measure
• MS : MSRA-TD50
• TT : Total-Text
• CT : CTW1500

Paper	year	IC13-R	IC13-P	IC13-F	IC15-R	IC15-P	IC15-F	MS-R	MS-P	MS-F	TT-R	TT-P	TT-F	CT-R	CT-P	CT-F
SOTA		95.0	97.4	0.925	91.8	0.9185	0.8984	91.5	88.2	81.7	88.3	87.6	84.7	86.9	86.0	83.4
[5]	16.04	0.78	0.88	0.83	0.43	0.71	0.54	0.67	0.83	0.74
[1] EAST	17.07	87.53	93.34	90.34	78.33	83.27	80.78	67.43	87.28	76.08	36.2	50.0	42.0	49.1	78.7	60.4
[7]	17.09	0.86	0.88	0.87	0.73	0.80	0.77
[3] PixelLink	18.01	88.6	87.5	88.1	82.0	85.5	83.7	83.0	73.2	77.8	54.41	59.89	57.02
[8] FOTS	18.01			0.925	0.8792	0.9185	0.8984
[9]	18.07	87.1	91.5	89.2				77.4	83.0	80.1
[10]	18.08	95.0	88.6	91.7	91.6	81.0	86.0				69.0	55.0	61.3
[6]	18.11	90.5	93.8	92.1	85.8	88.7	87.2				82.8	83.0	82.9
[4] MSR	19.01	88.5	91.8	90.1				76.7	87.4	81.7	73.0	85.2	78.6	77.8	83.8	80.7
[2]	19.01	85.94	93.18	89.41	79.2	86.1	82.5	75.26	85.88	80.21
[11]CRAFT	19.04	93.1	97.4		84.3	89.8		78.2	88.2		79.9	87.6		81.1	86.0
[12]					91.8	83.2	87.3	91.5	79.2	84.9	88.3	82.5	84.7	86.9	80.2	83.4

• [1] EAST: An Efficient and Accurate Scene Text Detector : [paper][review] : # U-Net
• [2] Detecting Text in the Wild with Deep Character Embedding Network
• [3] PixelLink: Detecting Scene Text via Instance Segmentation : [paper][review] : # U-Net
• [4] MSR: Multi-Scale Shape Regression for Scene Text Detection
• [5] Multi-Oriented Text Detection with Fully Convolutional Networks
• [6] Scene Text Detection with Supervised Pyramid Context Network : [paper][review]: #MASK R-CNN #Multi-Scale
• [7] Single Shot Text Detector with Regional Attention : #SSD
• [8] FOTS: Fast Oriented Text Spotting with a Unified Network : [paper][review]
• [9] Accurate Scene Text Detection through Border Semantics Awareness and Bootstrapping
• [10] Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes
• [11] CRAFT: Character Region Awareness for Text Detection, # F-Measure대신에 H-Mean으로 표현했다.같은것인지
• [12] Real-time Scene Text Detection with Differentiable Binarization

ocr's Issues

Text Detection 학습및 평가방법 고찰

학습셋

• SynthText dataset

Paper

[1] Accurate Scene Text Detection through Border Semantics Awareness and Bootstrapping # Multi-scale evaluation