Skip to content

Latest commit

 

History

History
164 lines (128 loc) · 10.3 KB

README.md

File metadata and controls

164 lines (128 loc) · 10.3 KB
Raw

«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 98.718 100,000 5,000
CRNN_Tiny CONV+GRU (1, 32, 160) 0.1 1.7 98.278 100,000 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 82.384 269,621 149,002
CRNN_Tiny CONV+GRU (3, 48, 168) 0.3 4.0 76.226 269,621 149,002

For each sub-dataset, the model performance as follows:

Model CCPD2019-Test Accuracy (%) Testing Data CCPD2020-Test Accuracy (%) Testing Data
CRNN 81.761 141,982 93.728 5,006
CRNN_Tiny 75.357 141,982 92.369 5,006

Table of Contents

News🚀🚀🚀

Version Release Date Major Updates
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.

Background

This warehouse aims to better understand and apply CRNN+CTC, and currently achieves digital recognition and license plate recognition

Installation

$ pip install -r requirements.txt

Usage

Train

# 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

Eval

# EMNIST
$ CUDA_VISIBLE_DEVICES=0 python eval_emnist.py crnn_tiny-emnist-b512-e100.pth ../datasets/emnist/
args: Namespace(not_tiny=False, pretrained='crnn_tiny-emnist-b512-e100.pth', use_lstm=False, val_root='../datasets/emnist/')
Loading CRNN pretrained: crnn_tiny-emnist-b512-e100.pth
crnn_tiny-emnist-b512-e100 summary: 22 layers, 427467 parameters, 427467 gradients, 0.1 GFLOPs
Batch:1562 ACC:100.000: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1563/1563 [00:19<00:00, 80.29it/s]
ACC:98.278
# Plate
$ CUDA_VISIBLE_DEVICES=0 python3 eval_plate.py crnn_tiny-plate-b512-e100.pth ../datasets/chinese_license_plate/recog/
args: Namespace(not_tiny=False, only_ccpd2019=False, only_ccpd2020=False, only_others=False, pretrained='crnn_tiny-plate-b512-e100.pth', use_lstm=False, val_root='../datasets/chinese_license_plate/recog/')
Loading CRNN pretrained: crnn_tiny-plate-b512-e100.pth
crnn_tiny-plate-b512-e100 summary: 22 layers, 1042318 parameters, 1042318 gradients, 0.3 GFLOPs
Load test data: 149002
Batch:4656 ACC:90.000: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4657/4657 [01:08<00:00, 67.50it/s]
ACC:76.226

Predict

$ CUDA_VISIBLE_DEVICES=0 python predict_emnist.py crnn_tiny-emnist-b512-e100.pth ../datasets/emnist/ ./runs/predict/emnist/
args: Namespace(not_tiny=False, pretrained='crnn_tiny-emnist-b512-e100.pth', save_dir='./runs/predict/emnist/', use_lstm=False, val_root='../datasets/emnist/')
Loading CRNN pretrained: crnn_tiny-emnist-b512-e100.pth
crnn_tiny-emnist-b512-e100 summary: 22 layers, 427467 parameters, 427467 gradients, 0.1 GFLOPs
Label: [3 8 5 8 5] Pred: [3 8 5 8 5]
Label: [4 8 6 8 0] Pred: [4 8 6 8 0]
Label: [4 6 4 7 0] Pred: [4 6 4 7 0]
Label: [2 3 5 0 7] Pred: [2 3 5 0 7]
Label: [4 7 8 4 6] Pred: [4 7 8 4 6]
Label: [0 1 4 3 6] Pred: [0 1 4 3 6]

$ CUDA_VISIBLE_DEVICES=0 python predict_plate.py crnn_tiny-plate-b512-e100.pth ./assets/plate/宁A87J92_0.jpg runs/predict/plate/
args: Namespace(image_path='./assets/plate/宁A87J92_0.jpg', not_tiny=False, pretrained='crnn_tiny-plate-b512-e100.pth', save_dir='runs/predict/plate/', use_lstm=False)
Loading CRNN pretrained: crnn_tiny-plate-b512-e100.pth
crnn_tiny-plate-b512-e100 summary: 22 layers, 1042318 parameters, 1042318 gradients, 0.3 GFLOPs
Pred: 宁A·87J92 - Predict time: 8.4 ms
Save to runs/predict/plate/plate_宁A87J92_0.jpg
$ CUDA_VISIBLE_DEVICES=0 python predict_plate.py crnn_tiny-plate-b512-e100.pth ./assets/plate/川A3X7J1_0.jpg runs/predict/plate/
args: Namespace(image_path='./assets/plate/川A3X7J1_0.jpg', not_tiny=False, pretrained='crnn_tiny-plate-b512-e100.pth', save_dir='runs/predict/plate/', use_lstm=False)
Loading CRNN pretrained: crnn_tiny-plate-b512-e100.pth
crnn_tiny-plate-b512-e100 summary: 22 layers, 1042318 parameters, 1042318 gradients, 0.3 GFLOPs
Pred: 川A·3X7J1 - Predict time: 8.4 ms
Save to runs/predict/plate/plate_川A3X7J1_0.jpg

Maintainers

  • zhujian - Initial work - zjykzj

Thanks

Contributing

Anyone's participation is welcome! Open an issue or submit PRs.

Small note:

License

Apache License 2.0 © 2023 zjykzj