LPSRGAN: License Plate Super-Resolution GAN

Unofficial implementation of the SR model proposed by Pan et al. in the paper "LPSRGAN: Generative Adversarial Networks for Super-Resolution of License Plate Image". This repository follows the methodology described by Nascimento et al. in "Toward Advancing License Plate Super-Resolution in Real-World Scenarios".

Prerequisites

Python 3.8+
PyTorch 1.10+
CUDA 11.3+ (for GPU acceleration)

Required packages: numpy, opencv-python, tqdm, yaml, torch Install via:

pip install -r requirements.txt

Installation

git clone https://github.com/your-username/LPSRGAN.git
cd LPSRGAN

Usage

This section provides instructions on testing the model, training it from scratch, and fine-tuning it on a custom dataset.

Testing

To test the model, ensure that the config file specifies the path to each of the .pth files (e.g., config file), as shown in the example below:

model:
  name: lpsrgan
  load: ./save/testing/model_lpsrgan.pth
  args:
    num_blocks: 3
    dropout_prob: 0.5

model_ocr:
  name: GPLPR
  load: ./save/testing/model_ocr_GPLPR.pth
  args:
    alphabet: "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
    nc: 3
    K: 7
    isSeqModel: True
    head: 2
    inner: 256
    isl2Norm: True

Once the configuration is set, execute the following command to start the test:

python3 test_mv.py --config ./config/testing.yaml --save ./results

Pretrained models are available at the following locations:
- Download GP_LPR from GP_LPR Unofficial Implementation.
- The pre-trained model for LPSRGAN is available in the releases section of this repository.

Training from Scratch

To train the model from scratch, update the following variables in the training config file:

resume: null  # Ensure no pretrained model is loaded
train_dataset:
  dataset:
    args:
      path_split: ./custom_split.txt

val_dataset:
  dataset:
    name: multi_image
    args:
      path_split: ./custom_split.txt
      phase: validation

Optionally, you can add the --tag argument for versioning:

python3 train.py --config ./config/training.yaml --save path/to/save/dest --tag experiment_1

Training on a Custom Dataset

To train or fine-tune the model on a custom dataset, you need to create a .txt file that lists the image folder paths along with their corresponding data split (training, validation, or testing). Each line in the file should follow this format:

<license_plate_text>;<image_path>;<split>

Example:

ABC1234;path/to/LP_image1.jpg;training
XYZ5678;path/to/LP_image2.jpg;validation

Update training.yaml to point to your split file:

resume: path/to/lpsrgan.pth  # Ensure the pretrained model is loaded

train_dataset:
  dataset:
    args:
      path_split: ./custom_split.txt

val_dataset:
  dataset:
    name: multi_image
    args:
      path_split: ./custom_split.txt
      phase: validation

For reference, you can check example files, such as split_all_example.txt (with its images and .txt files annotations located under dataset_intelbras_1280X960/mercosur and dataset_intelbras_1920x1080 directories), which demonstrate this format.

Configuration Files

Training: config/training.yaml
- Adjust hyperparameters, dataset paths, and optimizer settings here.
Testing: config/testing.yaml
- Specify model checkpoints and test dataset details.

Troubleshooting

CUDA Out of Memory: Reduce batch size in training.yaml.
Missing Pretrained Models: Ensure paths in testing.yaml or training.yaml are correct.
Dataset Errors: Validate split file formatting and image paths.

Citation

If you use this work, please cite the following works:

@article{pan2024lpsrgan,
  title={LPSRGAN: Generative adversarial networks for super-resolution of license plate image},
  author={Pan, Yuecheng and Tang, Jin and Tjahjadi, Tardi},
  journal={Neurocomputing},
  volume={580},
  pages={127426},
  year={2024},
  publisher={Elsevier}
}

@article{nascimento2025toward,
  title = {Toward Advancing License Plate Super-Resolution in Real-World Scenarios: A Dataset and Benchmark},
  author = {V. {Nascimento} and G. E. {Lima} and R. O. {Ribeiro} and W. R. {Schwartz} and R. {Laroca} and D. {Menotti}},
  year = {2025},
  journal = {Journal of the Brazilian Computer Society},
  volume = {},
  number = {},
  pages = {1-14},
  doi = {},
  issn = {},
}

Additionally, consider showing your support by starring :star: this repository.

Related Publications

Explore our other works on license plate recognition and super-resolution:

Contact

For questions or feedback, contact:

Valfride Wallace do Nascimento [Webpage]

vwnascimento@inf.ufpr.br

lpsrgan_for_test

Model card