News

• 2024-02-22: Updated for PyTorch 2.0 and Lightning 2.0
• 2024-01-16: Featured in the NVIDIA Developer Blog
• 2023-11-18: Interview with Deci AI at ECCV 2022 published
• 2023-09-07: Added to PaddleOCR, one of the most popular multilingual OCR toolkits
• 2023-06-15: Added to docTR, a deep learning-based library for OCR
• 2022-07-14: Initial public release (ranked #1 overall for STR on Papers With Code at the time of release)
• 2022-07-04: Accepted at ECCV 2022

Scene Text Recognition (STR) models use language context to be more robust against noisy or corrupted images. Recent approaches like ABINet use a standalone or external Language Model (LM) for prediction refinement. In this work, we show that the external LM—which requires upfront allocation of dedicated compute capacity—is inefficient for STR due to its poor performance vs cost characteristics. We propose a more efficient approach using permuted autoregressive sequence (PARSeq) models. View our ECCV poster and presentation for a brief overview.

NOTE: P-S and P-Ti are shorthands for PARSeq-S and PARSeq-Ti, respectively.

Method tl;dr

Our main insight is that with an ensemble of autoregressive (AR) models, we could unify the current STR decoding methods (context-aware AR and context-free non-AR) and the bidirectional (cloze) refinement model:

A single Transformer can realize different models by merely varying its attention mask. With the correct decoder parameterization, it can be trained with Permutation Language Modeling to enable inference for arbitrary output positions given arbitrary subsets of the input context. This arbitrary decoding characteristic results in a unified STR model—PARSeq—capable of context-free and context-aware inference, as well as iterative prediction refinement using bidirectional context without requiring a standalone language model. PARSeq can be considered an ensemble of AR models with shared architecture and weights:

NOTE: LayerNorm and Dropout layers are omitted. [B], [E], and [P] stand for beginning-of-sequence (BOS), end-of-sequence (EOS), and padding tokens, respectively. T = 25 results in 26 distinct position tokens. The position tokens both serve as query vectors and position embeddings for the input context. For [B], no position embedding is added. Attention masks are generated from the given permutations and are used only for the context-position attention. L_ce pertains to the cross-entropy loss.

Sample Results

Input Image	PARSeq-SA	ABINet	TRBA	ViTSTR-S	CRNN
	CHEWBACCA	CHEWBAGGA	CHEWBACCA	CHEWBACCA	CHEWUACCA
	Chevrol	Chevro_	Chevro_	Chevr__	Chevr__
	SALMON	SALMON	SALMON	SALMON	SA_MON
	Verbandsteffe	Verbandsteffe	Verbandstelle	Verbandsteffe	Verbandsleffe
	Kappa	Kappa	Kaspa	Kappa	Kaada
	3rdAve	3=-Ave	3rdAve	3rdAve	Coke

NOTE: Bold letters and underscores indicate wrong and missing character predictions, respectively.

Getting Started

This repository contains the reference implementation for PARSeq and reproduced models (collectively referred to as Scene Text Recognition Model Hub). See NOTICE for copyright information. Majority of the code is licensed under the Apache License v2.0 (see LICENSE) while ABINet and CRNN sources are released under the BSD and MIT licenses, respectively (see corresponding LICENSE files for details).

Demo

An interactive Gradio demo hosted at Hugging Face is available. The pretrained weights released here are used for the demo.

Installation

Requires Python >= 3.9 and PyTorch >= 2.0. The default requirements files will install the latest versions of the dependencies (as of February 22, 2024).

# Use specific platform build. Other PyTorch 2.0 options: cu118, cu121, rocm5.7
platform=cpu
# Generate requirements files for specified PyTorch platform
make torch-${platform}
# Install the project and core + train + test dependencies. Subsets: [dev,train,test,bench,tune]
pip install -r requirements/core.${platform}.txt -e .[train,test]

Updating dependency version pins

pip install pip-tools
make clean-reqs reqs  # Regenerate all the requirements files

Datasets

Download the datasets from the following links:

1. LMDB archives for MJSynth, SynthText, IIIT5k, SVT, SVTP, IC13, IC15, CUTE80, ArT, RCTW17, ReCTS, LSVT, MLT19, COCO-Text, and Uber-Text.
2. LMDB archives for TextOCR and OpenVINO.

Pretrained Models via Torch Hub

Available models are: abinet, crnn, trba, vitstr, parseq_tiny, parseq_patch16_224, and parseq.

import torch
from PIL import Image
from strhub.data.module import SceneTextDataModule

# Load model and image transforms
parseq = torch.hub.load('baudm/parseq', 'parseq', pretrained=True).eval()
img_transform = SceneTextDataModule.get_transform(parseq.hparams.img_size)

img = Image.open('/path/to/image.png').convert('RGB')
# Preprocess. Model expects a batch of images with shape: (B, C, H, W)
img = img_transform(img).unsqueeze(0)

logits = parseq(img)
logits.shape  # torch.Size([1, 26, 95]), 94 characters + [EOS] symbol

# Greedy decoding
pred = logits.softmax(-1)
label, confidence = parseq.tokenizer.decode(pred)
print('Decoded label = {}'.format(label[0]))

Frequently Asked Questions

• How do I train on a new language? See Issues #5 and #9.
• Can you export to TorchScript or ONNX? Yes, see Issue #12.
• How do I test on my own dataset? See Issue #27.
• How do I finetune and/or create a custom dataset? See Issue #7.
• What is val_NED? See Issue #10.

Training

The training script can train any supported model. You can override any configuration using the command line. Please refer to Hydra docs for more info about the syntax. Use ./train.py --help to see the default configuration.

./train.py +experiment=parseq-tiny pretrained=parseq-tiny  # Not all experiments have pretrained weights

./train.py +experiment=parseq-tiny  # Some examples: abinet-sv, trbc

./train.py charset=94_full  # Other options: 36_lowercase or 62_mixed-case. See configs/charset/

./train.py dataset=real  # Other option: synth. See configs/dataset/

./train.py model.img_size=[32, 128] model.max_label_length=25 model.batch_size=384

./train.py data.root_dir=data data.num_workers=2 data.augment=true

./train.py trainer.max_epochs=20 trainer.accelerator=gpu trainer.devices=2

./train.py +experiment=<model_exp> ckpt_path=outputs/<model>/<timestamp>/checkpoints/<checkpoint>.ckpt

Evaluation

The test script, test.py, can be used to evaluate any model trained with this project. For more info, see ./test.py --help.

PARSeq runtime parameters can be passed using the format param:type=value. For example, PARSeq NAR decoding can be invoked via ./test.py parseq.ckpt refine_iters:int=2 decode_ar:bool=false.

./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt  # or use the released weights: ./test.py pretrained=parseq

./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt  # lowercase alphanumeric (36-character set)
./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased  # mixed-case alphanumeric (62-character set)
./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased --punctuation  # mixed-case alphanumeric + punctuation (94-character set)

./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --new

./bench.py model=parseq model.decode_ar=false model.refine_iters=3
<torch.utils.benchmark.utils.common.Measurement object at 0x7f8fcae67ee0>
model(x)
  Median: 14.87 ms
  IQR:    0.33 ms (14.78 to 15.12)
  7 measurements, 10 runs per measurement, 1 thread
| module                | #parameters   | #flops   | #activations   |
|:----------------------|:--------------|:---------|:---------------|
| model                 | 23.833M       | 3.255G   | 8.214M         |
|  encoder              |  21.381M      |  2.88G   |  7.127M        |
|  decoder              |  2.368M       |  0.371G  |  1.078M        |
|  head                 |  36.575K      |  3.794M  |  9.88K         |
|  text_embed.embedding |  37.248K      |  0       |  0             |

./bench.py model=parseq model.decode_ar=false model.refine_iters=3 +range=true

./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased --punctuation  # no rotation
./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased --punctuation --rotation 90
./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased --punctuation --rotation 180
./test.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --cased --punctuation --rotation 270

./read.py outputs/<model>/<timestamp>/checkpoints/last.ckpt --images demo_images/*  # Or use ./read.py pretrained=parseq
Additional keyword arguments: {}
demo_images/art-01107.jpg: CHEWBACCA
demo_images/coco-1166773.jpg: Chevrol
demo_images/cute-184.jpg: SALMON
demo_images/ic13_word_256.png: Verbandsteffe
demo_images/ic15_word_26.png: Kaopa
demo_images/uber-27491.jpg: 3rdAve

# use NAR decoding + 2 refinement iterations for PARSeq
./read.py pretrained=parseq refine_iters:int=2 decode_ar:bool=false --images demo_images/*

Tuning

We use Ray Tune for automated parameter tuning of the learning rate. See ./tune.py --help. Extend tune.py to support tuning of other hyperparameters.

./tune.py tune.num_samples=20  # find optimum LR for PARSeq's default config using 20 trials
./tune.py +experiment=tune_abinet-lm  # find the optimum learning rate for ABINet's language model

Citation

@InProceedings{bautista2022parseq,
  title={Scene Text Recognition with Permuted Autoregressive Sequence Models},
  author={Bautista, Darwin and Atienza, Rowel},
  booktitle={European Conference on Computer Vision},
  pages={178--196},
  month={10},
  year={2022},
  publisher={Springer Nature Switzerland},
  address={Cham},
  doi={10.1007/978-3-031-19815-1_11},
  url={https://doi.org/10.1007/978-3-031-19815-1_11}
}