English | 中文
MASTER: Multi-Aspect Non-local Network for Scene Text Recognition
Attention-based scene text recognizers have gained huge success, which leverages a more compact intermediate representation to learn 1d- or 2d- attention by a RNN-based encoder-decoder architecture. However, such methods suffer from attention-drift problem because high similarity among encoded features leads to attention confusion under the RNN-based local attention mechanism. Moreover, RNN-based methods have low efficiency due to poor parallelization. To overcome these problems, this paper proposes the MASTER, a self-attention based scene text recognizer that (1) not only encodes the input-output attention but also learns self-attention which encodes feature-feature and target-target relationships inside the encoder and decoder and (2) learns a more powerful and robust intermediate representation to spatial distortion, and (3) owns a great training efficiency because of high training parallelization and a high-speed inference because of an efficient memory-cache mechanism. Extensive experiments on various benchmarks demonstrate the superior performance of MASTER on both regular and irregular scene text. [1]
Figure 1. Architecture of MASTER [1]
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 |
|---|---|---|---|---|---|---|
| Master-Resnet31 | D910x4-MS1.10-G | 90.37% | 6356 s/epoch | 2741 | 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Master-ResNet31 | 95.58% | 95.15% | 96.85% | 95.17% | 81.94% | 78.48% | 95.56% | 90.88% | 84.19% | 89.93% | 90.37% |
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 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 MASTER is (1, 3, 48, 160).
Please refer to the installation instruction in MindOCR.
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.zipcontains the datasets including training data, validation data and evaluation data.validation.zip: same as the validation/ within data_lmdb_release.zipevaluation.zip: same as the evaluation/ within data_lmdb_release.zip
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_fullthe ST_full contained the full cropped images of SynthText in LMDB data format. Please replace the ST folder with the ST_full folder.
Please download the SynthAdd Dataset from here (code: 627x). This dataset is proposed in https://arxiv.org/abs/1811.00751. Please prepare the corresponding LMDB dataset using the following command
python tools/dataset_converters/convert.py \
--dataset_name synthadd \
--task rec_lmdb \
--image_dir path_to_SynthAdd \
--output_path SynthAddPlease put the SynthAdd folder in /training directory.
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
│ └── SythAdd
│ ├── 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 17,402,659 samples)
- MJSynth (MJ)
- Train: 21.2 GB, 7224586 samples
- Valid: 2.36 GB, 802731 samples
- Test: 2.61 GB, 891924 samples
- SynthText Full (ST)
- 17.0 GB, 7266529 samples
- SynthAdd (SynthAdd)
- 2.7 GB, 1216889 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 scripttools/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
...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_sizeaccordingly 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.
- 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/master/master_resnet31.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/master/master_resnet31.yamlThe 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.
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/master/master_resnet31.yamlTo 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.
There are some built-in dictionaries, which are placed in mindocr/utils/dict/, and you can choose the appropriate dictionary to use.
en_dict.txtis an English dictionary containing 94 characters, including numbers, common symbols, and uppercase and lowercase English letters.ch_dict.txtis a Chinese dictionary containing 6623 characters, including commonly used simplified and traditional Chinese, numbers, common symbols, uppercase and lowercase English letters.
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_charin configuration yaml to True. - Remember to check the value of
dataset->transform_pipeline->RecMasterLabelEncode->lowerin the configuration yaml. Set it to False if you prefer case-sensitive encoding.
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 master_resnet31 --data_shape 48 160 --local_ckpt_path /path/to/local_ckpt.ckpt
# or
python tools/export.py --model_name_or_config configs/rec/master/master_resnet31.yaml --data_shape 48 160 --local_ckpt_path /path/to/local_ckpt.ckptThe 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/master/master_resnet31.yaml \
--res_save_dir=results_dir[1] Ning Lu, Wenwen Yu, Xianbiao Qi, Yihao Chen, Ping Gong, Rong Xiao, Xiang Bai. MASTER: Multi-Aspect Non-local Network for Scene Text Recognition. arXiv preprint arXiv:1910.02562, 2019.