English | 中文

SVTR

SVTR: Scene Text Recognition with a Single Visual Model

1. Introduction

Dominant scene text recognition models commonly contain two building blocks, a visual model for feature extraction and a sequence model for text transcription. This hybrid architecture, although accurate, is complex and less efficient. This paper proposes a Single Visual model for Scene Text recognition within the patch-wise image tokenization framework, which dispenses with the sequential modeling entirely. The method, termed SVTR, firstly decomposes an image text into small patches named character components. Afterward, hierarchical stages are recurrently carried out by component-level mixing, merging and/or combining. Global and local mixing blocks are devised to perceive the inter-character and intra-character patterns, leading to a multi-grained character component perception. Thus, characters are recognized by a simple linear prediction. Experimental results on both English and Chinese scene text recognition tasks demonstrate the effectiveness of SVTR. SVTR-L (Large) achieves highly competitive accuracy in English and outperforms existing methods by a large margin in Chinese, while running faster. In addition, SVTR-T (Tiny) is an effective and much smaller model, which shows appealing speed at inference. [1]

Figure 1. Architecture of SVTR [1]

2. Results

Accuracy

According to our experiments, the evaluation results on public benchmark datasets (IC03, IC13, IC15, IIIT, SVT, SVTP, CUTE) is as follow:

Model	Context	Avg Accuracy	Train T.	FPS	Recipe	Download
SVTR-Tiny	D910x4-MS1.10-G	90.23%	3638 s/epoch	4560	yaml	ckpt \| mindir

Detailed accuracy results for each benchmark dataset

Model	IC03_860	IC03_867	IC13_857	IC13_1015	IC15_1811	IC15_2077	IIIT5k_3000	SVT	SVTP	CUTE80	Average
SVTR-Tiny	95.70%	95.50%	95.33%	93.99%	83.60%	79.83%	94.70%	91.96%	85.58%	86.11%	90.23%

Notes:

Context: Training context denoted as {device}x{pieces}-{MS mode}, where mindspore mode can be G-graph mode or F-pynative mode with ms function. For example, D910x4-MS1.10-G is for training on 4 pieces of Ascend 910 NPU using graph mode based on Minspore version 1.10.
To reproduce the result on other contexts, please ensure the global batch size is the same.
The characters supported by model are lowercase English characters from a to z and numbers from 0 to 9. More explanation on dictionary, please refer to 4. Character Dictionary.
The models are trained from scratch without any pre-training. For more dataset details of training and evaluation, please refer to Dataset Download & Dataset Usage section.
The input Shapes of MindIR of RARE is (1, 3, 64, 256).

3. Quick Start

3.1 Preparation

3.1.1 Installation

Please refer to the installation instruction in MindOCR.

3.1.2 Dataset Preparation

3.1.2.1 MJSynth, validation and evaluation dataset

Part of the lmdb dataset for training and evaluation can be downloaded from here (ref: deep-text-recognition-benchmark). There're several zip files:

data_lmdb_release.zip contains the datasets including training data, validation data and evaluation data.
- training/ contains two datasets: MJSynth (MJ) and SynthText (ST). Here we use MJSynth only.
- validation/ is the union of the training sets of IC13, IC15, IIIT, and SVT.
- evaluation/ contains several benchmarking datasets, which are IIIT, SVT, IC03, IC13, IC15, SVTP, and CUTE.
validation.zip: same as the validation/ within data_lmdb_release.zip
evaluation.zip: same as the evaluation/ within data_lmdb_release.zip

3.1.2.2 SynthText dataset

For SynthText, we do not use the given LMDB dataset in data_lmdb_release.zip, since it only contains part of the cropped images. Please download the raw dataset from here and prepare the LMDB dataset using the following command

python tools/dataset_converters/convert.py \
    --dataset_name synthtext \
    --task rec_lmdb \
    --image_dir path_to_SynthText \
    --label_dir path_to_SynthText_gt.mat \
    --output_path ST_full

the ST_full contained the full cropped images of SynthText in LMDB data format. Please replace the ST folder with the ST_full folder.

3.1.3 Dataset Usage

Finally, the data structure should like this.

data_lmdb_release/
├── evaluation
│   ├── CUTE80
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC03_860
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC03_867
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC13_1015
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── ...
├── training
│   ├── MJ
│   │   ├── MJ_test
│   │   │   ├── data.mdb
│   │   │   └── lock.mdb
│   │   ├── MJ_train
│   │   │   ├── data.mdb
│   │   │   └── lock.mdb
│   │   └── MJ_valid
│   │       ├── data.mdb
│   │       └── lock.mdb
│   └── ST_full
│       ├── data.mdb
│       └── lock.mdb
└── validation
    ├── data.mdb
    └── lock.mdb

Here we used the datasets under training/ folders for training, and the union dataset validation/ for validation. After training, we used the datasets under evaluation/ to evaluate model accuracy.

Training: (total 16,185,770 samples)

MJSynth (MJ)
- Train: 21.2 GB, 7224586 samples
- Valid: 2.36 GB, 802731 samples
- Test: 2.61 GB, 891924 samples
SynthText (ST)
- Train: 17.0 GB, 7266529 samples

Validation:

Valid: 138 MB, 6992 samples

Evaluation: (total 12,067 samples)

CUTE80: 8.8 MB, 288 samples
IC03_860: 36 MB, 860 samples
IC03_867: 4.9 MB, 867 samples
IC13_857: 72 MB, 857 samples
IC13_1015: 77 MB, 1015 samples
IC15_1811: 21 MB, 1811 samples
IC15_2077: 25 MB, 2077 samples
IIIT5k_3000: 50 MB, 3000 samples
SVT: 2.4 MB, 647 samples
SVTP: 1.8 MB, 645 samples

Data configuration for model training

To reproduce the training of model, it is recommended that you modify the configuration yaml as follows:

...
train:
  ...
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of training dataset
    data_dir: training/                                               # Dir of training dataset, concatenated with `dataset_root` to be the complete dir of training dataset
...
eval:
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of validation dataset
    data_dir: validation/                                             # Dir of validation dataset, concatenated with `dataset_root` to be the complete dir of validation dataset
  ...

Data configuration for model evaluation

We use the dataset under evaluation/ as the benchmark dataset. On each individual dataset (e.g. CUTE80, IC03_860, etc.), we perform a full evaluation by setting the dataset's directory to the evaluation dataset. This way, we get a list of the corresponding accuracies for each dataset, and then the reported accuracies are the average of these values.

To reproduce the reported evaluation results, you can:

Option 1: Repeat the evaluation step for all individual datasets: CUTE80, IC03_860, IC03_867, IC13_857, IC131015, IC15_1811, IC15_2077, IIIT5k_3000, SVT, SVTP. Then take the average score.
Option 2: Put all the benchmark datasets folder under the same directory, e.g. evaluation/. And use the script tools/benchmarking/multi_dataset_eval.py.

Evaluate on one specific dataset

For example, you can evaluate the model on dataset CUTE80 by modifying the config yaml as follows:

...
train:
  # NO NEED TO CHANGE ANYTHING IN TRAIN SINCE IT IS NOT USED
...
eval:
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of evaluation dataset
    data_dir: evaluation/CUTE80/                                      # Dir of evaluation dataset, concatenated with `dataset_root` to be the complete dir of evaluation dataset
  ...

By running tools/eval.py as noted in section Model Evaluation with the above config yaml, you can get the accuracy performance on dataset CUTE80.

Evaluate on multiple datasets under the same folder

Assume you have put all benckmark datasets under evaluation/ as shown below:

data_lmdb_release/
├── evaluation
│   ├── CUTE80
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC03_860
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC03_867
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── IC13_1015
│   │   ├── data.mdb
│   │   └── lock.mdb
│   ├── ...

then you can evaluate on each dataset by modifying the config yaml as follows, and execute the script tools/benchmarking/multi_dataset_eval.py.

...
train:
  # NO NEED TO CHANGE ANYTHING IN TRAIN SINCE IT IS NOT USED
...
eval:
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of evaluation dataset
    data_dir: evaluation/                                   # Dir of evaluation dataset, concatenated with `dataset_root` to be the complete dir of evaluation dataset
  ...

3.1.4 Check YAML Config Files

Apart from the dataset setting, please also check the following important args: system.distribute, system.val_while_train, common.batch_size, train.ckpt_save_dir, train.dataset.dataset_root, train.dataset.data_dir, train.dataset.label_file, eval.ckpt_load_path, eval.dataset.dataset_root, eval.dataset.data_dir, eval.dataset.label_file, eval.loader.batch_size. Explanations of these important args:

system:
  distribute: True                                                    # `True` for distributed training, `False` for standalone training
  amp_level: 'O2'
  amp_level_infer: "O2"
  seed: 42
  val_while_train: True                                               # Validate while training
  drop_overflow_update: False
common:
  ...
  batch_size: &batch_size 512                                         # Batch size for training
...
train:
  ckpt_save_dir: './tmp_rec'                                          # The training result (including checkpoints, per-epoch performance and curves) saving directory
  dataset_sink_mode: False
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of training dataset
    data_dir: training/                                               # Dir of training dataset, concatenated with `dataset_root` to be the complete dir of training dataset
...
eval:
  ckpt_load_path: './tmp_rec/best.ckpt'                               # checkpoint file path
  dataset_sink_mode: False
  dataset:
    type: LMDBDataset
    dataset_root: dir/to/data_lmdb_release/                           # Root dir of validation/evaluation dataset
    data_dir: validation/                                             # Dir of validation/evaluation dataset, concatenated with `dataset_root` to be the complete dir of validation/evaluation dataset
  ...
  loader:
      shuffle: False
      batch_size: 512                                                 # Batch size for validation/evaluation
...

Notes:

As the global batch size (batch_size x num_devices) is important for reproducing the result, please adjust batch_size accordingly to keep the global batch size unchanged for a different number of GPUs/NPUs, or adjust the learning rate linearly to a new global batch size.

3.2 Model Training

Distributed Training

It is easy to reproduce the reported results with the pre-defined training recipe. For distributed training on multiple Ascend 910 devices, please modify the configuration parameter distribute as True and run

# distributed training on multiple GPU/Ascend devices
mpirun --allow-run-as-root -n 4 python tools/train.py --config configs/rec/svtr/svtr_tiny.yaml

Standalone Training

If you want to train or finetune the model on a smaller dataset without distributed training, please modify the configuration parameterdistribute as False and run:

# standalone training on a CPU/GPU/Ascend device
python tools/train.py --config configs/rec/svtr/svtr_tiny.yaml

The training result (including checkpoints, per-epoch performance and curves) will be saved in the directory parsed by the arg ckpt_save_dir. The default directory is ./tmp_rec.

3.3 Model Evaluation

To evaluate the accuracy of the trained model, you can use eval.py. Please set the checkpoint path to the arg ckpt_load_path in the eval section of yaml config file, set distribute to be False, and then run:

python tools/eval.py --config configs/rec/svtr/svtr_tiny.yaml

4. Character Dictionary

Default Setting

To transform the groud-truth text into label ids, we have to provide the character dictionary where keys are characters and values are IDs. By default, the dictionary is "0123456789abcdefghijklmnopqrstuvwxyz", which means id=0 will correspond to the charater "0". In this case, the dictionary only considers numbers and lowercase English characters, excluding spaces.

Built-in Dictionaries

There are some built-in dictionaries, which are placed in mindocr/utils/dict/, and you can choose the appropriate dictionary to use.

en_dict.txt is an English dictionary containing 94 characters, including numbers, common symbols, and uppercase and lowercase English letters.
ch_dict.txt is a Chinese dictionary containing 6623 characters, including commonly used simplified and traditional Chinese, numbers, common symbols, uppercase and lowercase English letters.

Customized Dictionary

You can also customize a dictionary file (***.txt) and place it under mindocr/utils/dict/, the format of the dictionary file should be a .txt file with one character per line.

To use a specific dictionary, set the parameter character_dict_path to the path of the dictionary, and change the parameter num_classes to the corresponding number, which is the number of characters in the dictionary + 1.

Notes:

You can include the space character by setting the parameter use_space_char in configuration yaml to True.
Remember to check the value of dataset->transform_pipeline->RecAttnLabelEncode->lower in the configuration yaml. Set it to False if you prefer case-sensitive encoding.

5. Chinese Text Recognition Model Training

Currently, this model supports multilingual recognition and provides pre-trained models for different languages. Details are as follows:

Chinese Dataset Preparation and Configuration

We use a public Chinese text benchmark dataset Benchmarking-Chinese-Text-Recognition for SVTR training and evaluation.

For detailed instruction of data preparation and yaml configuration, please refer to ch_dataeset.

Training

To train with the prepared datsets and config file, please run:

mpirun --allow-run-as-root -n 4 python tools/train.py --config configs/rec/svtr/svtr_tiny_ch.yaml

Results and Pretrained Weights

After training, evaluation results on the benchmark test set are as follows, where we also provide the model config and pretrained weights.

Model	Language	Context	Scene	Web	Document	Train T.	FPS	Recipe	Download
SVTR-Tiny	Chinese	D910x4-MS1.10-G	65.93%	69.64%	98.01%	647 s/epoch	1580	svtr_tiny_ch.yaml	ckpt \| mindir

Training with Custom Datasets

You can train models for different languages with your own custom datasets. Loading the pretrained Chinese model to finetune on your own dataset usually yields better results than training from scratch. Please refer to the tutorial Training Recognition Network with Custom Datasets.

6. MindSpore Lite Inference

To inference with MindSpot Lite on Ascend 310, please refer to the tutorial MindOCR Inference. In short, the whole process consists of the following steps:

1. Model Export

Please download the exported MindIR file first, or refer to the Model Export tutorial and use the following command to export the trained ckpt model to MindIR file:

python tools/export.py --model_name_or_config svtr_tiny --data_shape 64 256 --local_ckpt_path /path/to/local_ckpt.ckpt
# or
python tools/export.py --model_name_or_config configs/rec/svtr/svtr_tiny.yaml --data_shape 64 256 --local_ckpt_path /path/to/local_ckpt.ckpt

The data_shape is the model input shape of height and width for MindIR file. The shape value of MindIR in the download link can be found in Notes under results table.

2. Environment Installation

Please refer to Environment Installation tutorial to configure the MindSpore Lite inference environment.

3. Model Conversion

Please refer to Model Conversion, and use the converter_lite tool for offline conversion of the MindIR file.

4. Inference

Assuming that you obtain output.mindir after model conversion, go to the deploy/py_infer directory, and use the following command for inference:

python infer.py \
    --input_images_dir=/your_path_to/test_images \
    --rec_model_path=your_path_to/output.mindir \
    --rec_model_name_or_config=../../configs/rec/svtr/svtr_tiny.yaml \
    --res_save_dir=results_dir

References

[1] Yongkun Du, Zhineng Chen, Caiyan Jia, Xiaoting Yin, Tianlun Zheng, Chenxia Li, Yuning Du, Yu-Gang Jiang. SVTR: Scene Text Recognition with a Single Visual Model. arXiv preprint arXiv:2205.00159, 2022.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

SVTR

1. Introduction

2. Results

Accuracy

3. Quick Start

3.1 Preparation

3.1.1 Installation

3.1.2 Dataset Preparation

3.1.2.1 MJSynth, validation and evaluation dataset

3.1.2.2 SynthText dataset

3.1.3 Dataset Usage

3.1.4 Check YAML Config Files

3.2 Model Training

3.3 Model Evaluation

4. Character Dictionary

Default Setting

Built-in Dictionaries

Customized Dictionary

5. Chinese Text Recognition Model Training

Chinese Dataset Preparation and Configuration

Training

Results and Pretrained Weights

Training with Custom Datasets

6. MindSpore Lite Inference

References

Files

README.md

Latest commit

History

README.md

File metadata and controls

SVTR

1. Introduction

2. Results

Accuracy

3. Quick Start

3.1 Preparation

3.1.1 Installation

3.1.2 Dataset Preparation

3.1.2.1 MJSynth, validation and evaluation dataset

3.1.2.2 SynthText dataset

3.1.3 Dataset Usage

3.1.4 Check YAML Config Files

3.2 Model Training

3.3 Model Evaluation

4. Character Dictionary

Default Setting

Built-in Dictionaries

Customized Dictionary

5. Chinese Text Recognition Model Training

Chinese Dataset Preparation and Configuration

Training

Results and Pretrained Weights

Training with Custom Datasets

6. MindSpore Lite Inference

References