«crnn-ctc» implemented CRNN+CTC
ONLINE DEMO：LICENSE PLATE RECOGNITION
| Model | ARCH | Input Shape | GFLOPs | Model Size (MB) | EMNIST Accuracy (%) | Training Data | Testing Data |
|---|---|---|---|---|---|---|---|
| CRNN | CONV+GRU | (1, 32, 160) | 2.2 | 31 | 98.570 | 100,000 | 5,000 |
| CRNN_Tiny | CONV+GRU | (1, 32, 160) | 0.1 | 1.7 | 98.306 | 100,000 | 5,000 |
| Model | ARCH | Input Shape | GFLOPs | Model Size (MB) | ChineseLicensePlate Accuracy (%) | Training Data | Testing Data |
|---|---|---|---|---|---|---|---|
| CRNN | CONV+GRU | (3, 48, 168) | 4.0 | 58 | 82.147 | 269,621 | 149,002 |
| CRNN_Tiny | CONV+GRU | (3, 48, 168) | 0.3 | 4.0 | 76.590 | 269,621 | 149,002 |
| LPRNetPlus | CONV | (3, 24, 94) | 0.5 | 2.3 | 63.546 | 269,621 | 149,002 |
| LPRNet | CONV | (3, 24, 94) | 0.3 | 1.9 | 60.105 | 269,621 | 149,002 |
| LPRNetPlus+STNet | CONV | (3, 24, 94) | 0.5 | 2.5 | 72.130 | 269,621 | 149,002 |
| LPRNet+STNet | CONV | (3, 24, 94) | 0.3 | 2.2 | 72.261 | 269,621 | 149,002 |
For each sub-dataset, the model performance as follows:
| Model | CCPD2019-Test Accuracy (%) | Testing Data | CCPD2020-Test Accuracy (%) | Testing Data |
|---|---|---|---|---|
| CRNN | 81.512 | 141,982 | 93.787 | 5,006 |
| CRNN_Tiny | 75.729 | 141,982 | 92.829 | 5,006 |
| LPRNetPlus | 62.184 | 141,982 | 89.373 | 5,006 |
| LPRNet | 59.597 | 141,982 | 89.153 | 5,006 |
| LPRNetPlus+STNet | 72.125 | 141,982 | 90.611 | 5,006 |
| LPRNet+STNet | 71.291 | 141,982 | 89.832 | 5,006 |
| Version | Release Date | Major Updates |
|---|---|---|
| v1.3.0 | 2024/09/21 | Add STNet module to LPRNet/LPRNetPlus and update the training/evaluation/prediction results on the CCPD dataset. |
| v1.2.0 | 2024/09/17 | Create a new LPRNet/LPRNetPlus model and update the training/evaluation/prediction results on the CCPD dataset. |
| v1.1.0 | 2024/08/17 | Update EVAL/PREDICT implementation, support Pytorch format model conversion to ONNX, and finally provide online demo based on Gradio. |
| v1.0.0 | 2024/08/04 | Optimize the CRNN architecture while achieving super lightweight CRNN_Tiny. In addition, all training scripts support mixed precision training. |
| v0.3.0 | 2024/08/03 | Implement models CRNN_LSTM and CRNN_GRU on datasets EMNIST and ChineseLicensePlate. |
| v0.2.0 | 2023/10/11 | Support training/evaluation/prediction of CRNN+CTC based on license plate. |
| v0.1.0 | 2023/10/10 | Support training/evaluation/prediction of CRNN+CTC based on EMNIST digital characters. |
This warehouse aims to better understand and apply CRNN+CTC, and has currently achieved digital recognition and license plate recognition. Meanwhile, LPRNet(+STNet) is a pure convolutional architecture for license plate recognition network. I believe that the implementation of these algorithms can help with the deployment of license plate recognition algorithms, such as on edge devices.
Relevant papers include:
- Towards End-to-End License Plate Detection and Recognition: A Large Dataset and Baseline
- An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
- Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks
- LPRNet: License Plate Recognition via Deep Neural Networks
Relevant blogs (Chinese):
- Towards End-to-End License Plate Detection and Recognition: A Large Dataset and Baseline
- An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
- LPRNet: License Plate Recognition via Deep Neural Networks
$ pip install -r requirements.txtOr use docker container
$ docker run -it --runtime nvidia --gpus=all --shm-size=16g -v /etc/localtime:/etc/localtime -v $(pwd):/workdir --workdir=/workdir --name crnn-ctc ultralytics/yolov5:latest- ChineseLicensePlate: Baidu Drive(ad7l)
# EMNIST
$ python3 train_emnist.py ../datasets/emnist/ ./runs/crnn-emnist-b512/ --batch-size 512 --device 0 --not-tiny
# Plate
$ python3 train_plate.py ../datasets/chinese_license_plate/recog/ ./runs/crnn-plate-b512/ --batch-size 512 --device 0 --not-tiny# EMNIST
$ CUDA_VISIBLE_DEVICES=0 python eval_emnist.py crnn-emnist.pth ../datasets/emnist/ --not-tiny
args: Namespace(not_tiny=True, pretrained='crnn-emnist.pth', use_lstm=False, val_root='../datasets/emnist/')
Loading CRNN pretrained: crnn-emnist.pth
crnn-emnist summary: 29 layers, 7924363 parameters, 7924363 gradients, 2.2 GFLOPs
Batch:49999 ACC:100.000: 100%|████████████████████████████████████████████████████████| 50000/50000 [03:47<00:00, 219.75it/s]
ACC:98.570
# Plate
$ CUDA_VISIBLE_DEVICES=0 python3 eval_plate.py crnn-plate.pth ../datasets/chinese_license_plate/recog/ --not-tiny
args: Namespace(add_stnet=False, not_tiny=True, only_ccpd2019=False, only_ccpd2020=False, only_others=False, pretrained='crnn-plate.pth', use_lprnet=False, use_lstm=False, use_origin_block=False, val_root='../datasets/chinese_license_plate/recog/')
Loading CRNN pretrained: crnn-plate.pth
crnn-plate summary: 29 layers, 15083854 parameters, 15083854 gradients, 4.0 GFLOPs
Load test data: 149002
Batch:4656 ACC:100.000: 100%|████████████████████████████████████████████████████████████| 4657/4657 [00:52<00:00, 89.13it/s]
ACC:82.147$ CUDA_VISIBLE_DEVICES=0 python predict_emnist.py crnn-emnist.pth ../datasets/emnist/ ./runs/predict/emnist/ --not-tiny
args: Namespace(not_tiny=True, pretrained='crnn-emnist.pth', save_dir='./runs/predict/emnist/', use_lstm=False, val_root='../datasets/emnist/')
Loading CRNN pretrained: crnn-emnist.pth
crnn-emnist summary: 29 layers, 7924363 parameters, 7924363 gradients, 2.2 GFLOPs
Label: [0 4 2 4 7] Pred: [0 4 2 4 7]
Label: [2 0 6 5 4] Pred: [2 0 6 5 4]
Label: [7 3 9 9 5] Pred: [7 3 9 9 5]
Label: [9 6 6 0 9] Pred: [9 6 6 0 9]
Label: [2 3 0 7 6] Pred: [2 3 0 7 6]
Label: [6 5 9 5 2] Pred: [6 5 9 5 2]
$ CUDA_VISIBLE_DEVICES=0 python predict_plate.py crnn-plate.pth ./assets/plate/宁A87J92_0.jpg runs/predict/plate/ --not-tiny
args: Namespace(add_stnet=False, image_path='./assets/plate/宁A87J92_0.jpg', not_tiny=True, pretrained='crnn-plate.pth', save_dir='runs/predict/plate/', use_lprnet=False, use_lstm=False, use_origin_block=False)
Loading CRNN pretrained: crnn-plate.pth
crnn-plate summary: 29 layers, 15083854 parameters, 15083854 gradients, 4.0 GFLOPs
Pred: 宁A·87J92 - Predict time: 5.4 ms
Save to runs/predict/plate/plate_宁A87J92_0.jpg
$ CUDA_VISIBLE_DEVICES=0 python predict_plate.py crnn-plate.pth ./assets/plate/川A3X7J1_0.jpg runs/predict/plate/ --not-tiny
args: Namespace(add_stnet=False, image_path='./assets/plate/川A3X7J1_0.jpg', not_tiny=True, pretrained='crnn-plate.pth', save_dir='runs/predict/plate/', use_lprnet=False, use_lstm=False, use_origin_block=False)
Loading CRNN pretrained: crnn-plate.pth
crnn-plate summary: 29 layers, 15083854 parameters, 15083854 gradients, 4.0 GFLOPs
Pred: 川A·3X7J1 - Predict time: 4.7 ms
Save to runs/predict/plate/plate_川A3X7J1_0.jpg
- zhujian - Initial work - zjykzj
- rinabuoy/crnn-ctc-loss-pytorch
- we0091234/crnn_plate_recognition
- sirius-ai/LPRNet_Pytorch
- CA-USTC/License_Plate_Recognition_pytorch
- zjykzj/LPDet
Anyone's participation is welcome! Open an issue or submit PRs.
Small note:
- Git submission specifications should be complied with Conventional Commits
- If versioned, please conform to the Semantic Versioning 2.0.0 specification
- If editing the README, please conform to the standard-readme specification.
Apache License 2.0 © 2023 zjykzj