GP_LPR

Unofficial implementation of the OCR model proposed by Liu et al. in the paper "Irregular License Plate Recognition via Global Information Integration."

Dataset

The RodoSol-ALPR dataset is a valuable resource for research in Automatic License Plate Recognition (ALPR). It contains 20,000 images of various vehicle types, including cars and motorcycles, captured under different conditions — day and night, across multiple toll stations, and during both clear and rainy weather.

Usage

This section provides instructions on testing the model, training it from scratch, and fine-tuning it on a custom dataset. Follow the steps below to set up and run the model. Additionally,

Testing

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

model_ocr:
  name: GPLPR
  load: ./save/testing/model.pth
  args:
    nc: 3
    isSeqModel: True
    head: 2
    inner: 256
    isl2Norm: True

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

python3 test_ocr.py --config ./config/testing.yaml --save True --tag example

Training from Scratch

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

resume: null

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

python3 train.py --config ./config/training.yaml --save True

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

python3 train.py --config ./config/training.yaml --save True --tag example

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 paths along with their corresponding data split (training, validation, or testing). Each line in the file should follow this format:

path/to/LP_image1.jpg;training
path/to/LP_image2.jpg;validation
path/to/LP_image3.jpg;testing

For reference, you can check example files, such as train_dir_split.txt (with its images and .txt files annotations located under train_dir directory), split_all_pku.txt and split_all_rodosol.txt, which demonstrate this format.

Modifying the Configuration File for Training/Finetuning

To customize the model for training or fine-tuning on a custom dataset, follow these steps:

1. Create a custom alphabet: Update all fields marked with alphabet: "customAlphabet" in the config file to reflect your specific alphabet.
2. Update the dataset split: Modify the path_split argument to point to your custom dataset split file (e.g., your_custom_split.txt).

Here’s an example of the modified training.yaml file:

model:
  name: GPLPR
  OCR_TRAIN: True
  args:
    nc: 3
    alphabet: "customAlphabet"   # Specify your custom alphabet here
    K: 7
    isSeqModel: True
    head: 2
    inner: 256
    isl2Norm: True

func_train: GP_LPR_TRAIN
func_val: GP_LPR_VAL
alphabet: "customAlphabet"         # Apply the custom alphabet

train_dataset:
  dataset:
    name: ocr_img
    args:
      path_split: your_custom_split.txt  # Set the path to your custom split file
      phase: training

  wrapper:
    name: Ocr_images_lp
    args:
      alphabet: "customAlphabet"   # Specify your custom alphabet here
      k: 7
      imgW: 96
      imgH: 32
      aug: True
      image_aspect_ratio: 3
      background: (127, 127, 127)
      with_lr: False
  batch: 128

val_dataset:
  dataset:
    name: ocr_img
    args:
      path_split: your_custom_split.txt  # Set the path to your custom split file
      phase: validation

  wrapper:
    name: Ocr_images_lp
    args:
      alphabet: "customAlphabet"   # Specify your custom alphabet here
      k: 7
      imgW: 96
      imgH: 32
      aug: False
      image_aspect_ratio: 3
      background: (127, 127, 127)
      with_lr: False
  batch: 128

optimizer:
  name: adam
  args:
    lr: 1.e-3
    betas: [0.5, 0.555]

epoch_max: 3000

loss:
  name: CrossEntropyLoss
  args:
    size_average: None
    reduce: None
    reduction: mean

early_stopper:
  patience: 400
  min_delta: 0
  counter: 0

epoch_max: 3000
epoch_save: 100
resume: null   # For fine-tuning, point this to your pre-trained model path

Additional Notes:

• Fine-tuning: If you’re fine-tuning the model, modify the resume: field to point to your pre-trained model file.
• Alphabet: Make sure to specify the correct custom alphabet consistently in all sections.
• Dataset split: Ensure the your_custom_split.txt file contains the correct paths and data splits (e.g., training, validation, testing).