A super-resolution framework for license plate recognition

Project

Setting up

• please create a conda environment
• install the packages according to requirements.txt
• please clone the project and download the proposed dataset in google drive (HR_782.zip, LR_782.zip)

Training

• define the .yml file under ./options/train/SwinFIR
  • name of the experiment
  • dataset directory
  • scheduler, total_iter
  • loss functions
• run the command (e.g. below)

CUDA_VISIBLE_DEVICES=0 python swinfir/train.py -opt options/train/SwinFIR/train_crnn4_Relu_VGG.yml

• models and training_states will be saved under ./experiments

Tensorboard

• watch the PSNR/SSIM curves and loss value curves during validation

tensorboard --logdir tb_logger/crnnOCR_continue --port 5500 --bind_all

Testing

• modify the .yml file ./options/test/SwinFIR/SwinFIR_SRx2.yml
  • name of testing
  • dataset directory
  • model path (pretrain_network_g)
• run the command (e.g. below)

python swinfir/test.py -opt options/test/SwinFIR/SwinFIR_SRx2.yml

• the output super-resolution images will be saved under ./results

OCR recognizers

Multi-task

• the OCR model is download from https://github.com/Valfride/lpr-rsr-ext
• please refer to PKU-SR.zip in google drive
• run OCRpred_SR.py to get the recognition results from Multi-task

PaddleOCR

• please clone the project https://github.com/PaddlePaddle/PaddleOCR
• please put model file under ./models
  • please refer to ch_PP-OCRv3_rec_train.zip in google drive
• modify /configs/ch_PP-OCRv3_rec.yml
  • change infer_img
• modify /tools/infer_rec.py
  • change output_file_path
• run the command

python3 tools/infer_rec.py -c configs/ch_PP-OCRv3_rec.yml

• run calculate_acc.py to output .csv file
  • change the input txt path and output filename

CRNN

• please clone the project from https://github.com/we0091234/crnn_plate_recognition/tree/master
• the model is already under ./saved_model
• please modify ./demo.py
  • change image_path and output text file path
• run calculate_acc.py to output .csv file
  • change the input txt path and output filename

Dataset

Proposed dataset (in google drive)

• HR_782 and LR_782 are used for our project
• LR_bicubic is for showing LR image with the same size of HR image, it has been resized using bicubic method to x2
• HR_ori_size and LR_ori_size : each image is in their original size
• LSVLP_cropped_beforePS : each image is unrectified by Photoshop

PKU-SR dataset (in google drive)

• original PKU-SR dataset is PKU-Dataset-SR.zip
  • clone the project https://github.com/Valfride/lpr-rsr-ext to train and test on it
• PKUSR.zip is for training and testing our method
  • data has been split to train/val/test folder

Others (in google drive)

• FSI-DI-Dataset is the original dataset from the paper https://www.sciencedirect.com/science/article/pii/S2666281720303899
• FSI-DI : HR-LR paired images which are collected from FSI-DI-Dataset