method: Tencent TEG OCR2020-03-15
Authors: Pei Xu, Shan Huang, Hongzhen Wang, Shen Huang, Qi Ju.
Description: We reimplemented the standalone recognition method according to the end-to-end text spotting code released by the Mask TextSpotter[TPAMI]. It is a seq-to-seq method based on 2D attention. We synthesize curved text images for pretraining by the method of VGG synthtext. We add public dataset including icdar2013-2015, CUTE, SVT, IIIT5k, RCTW2017, LSVT to finetune and don't use any private data.
method: baseline_0.5_class_54352019-04-28
Authors: Jinjin Zhang, Beihang University
Description: instance segment based method for text detection and attention based method for text recognition with threshold 0.5 and 5435 classes. Data augmentation and extra datasets including LSVT, ICDAR2017, COCO-Text, RECTS are used for text recognition.
method: Detection-Recognition2019-04-30
Authors: USTC-iFLYTEK
Description: Two-stage detection-recognition Text Spotting: We just combine our text detection model and text line recognition model. For each detect result represented as a polygon, we crop a sub text line image with minimum bounding rectangle box as input image of our text line recognition model.
| Date | Method | Recall | Precision | Hmean | 1-NED | |||
|---|---|---|---|---|---|---|---|---|
| 2020-03-15 | Tencent TEG OCR | 66.61% | 70.30% | 68.40% | 67.12% | |||
| 2019-04-28 | baseline_0.5_class_5435 | 49.29% | 56.03% | 52.45% | 53.86% | |||
| 2019-04-30 | Detection-Recognition | 40.77% | 60.29% | 48.64% | 45.84% | |||
| 2019-04-29 | task3 | 30.03% | 49.84% | 37.48% | 34.03% | |||
| 2019-04-30 | CRAFT + TPS-ResNet v1 | 26.43% | 39.53% | 31.68% | 27.21% |