The repository contains scripts required for OCR Pipeline Workflow on Onepanel. This pipeline consist of two stages-- license plate detection and license plate ocr.

For the detection, the object detection model was trained from CVAT. The default Workflow also exports frozen graph which can be used in license_detection.py to generate the output.

Once we have the output from detection model, it can fed to the Attention OCR model for OCR. Following steps demonstrate the training and inference process for this model.

1 - Generate TFRecords

python generate_tfrecords.py \
        --charset_path=/mnt/src/train/data/charset_size.txt \
        --data_dir=/mnt/data/datasets/ \
        --output_tfrecord=/mnt/data/datasets/tfexample_train  \
        --text_length=20

Upon some data exploration, it was found that the maximum length of text in our dataset was 31. So, the model was trained with text_length set to 31. But the resulting model was very inaccurate. With some more experiments, we found that the number of examples with the length of text more than 20 were 10. In other words, for most of the samples there were more null characters than the alphabets.

2 - Train

If you want to fine tune:

wget http://download.tensorflow.org/models/attention_ocr_2017_08_09.tar.gz && \
tar xf attention_ocr_2017_08_09.tar.gz 

Export path for our custom.py dataset script:

export PYTHONPATH=$PYTHONPATH:./datasets/

python train.py \
    --dataset_name=custom \
    --batch_size=1 \
    --train_log_dir=/mnt/output/ \
    --checkpoint=model.ckpt-399731
    --max_number_of_steps=110000

3 - Run inference

python3 demo_inference.py --dataset_name=custom --batch_size=1 --image_path_pattern=/data/trd/ --checkpoint=/data/model.ckpt-110000

4 - Evaluation

python3 eval.py --dataset_name=custom --batch_size=1 --split_name=test --num_batches=9 --train_log_dir=/data/