Convolutional Recurrent Neural Network + CTCLoss | STAR-Net

Code for paper "Towards Boosting the Accuracy of Non-Latin Scene Text Recognition"

Dependence

• Python3.6.5
• torch==1.2.0
• torchvision==0.4.0
• tensorboard==2.3.0

How to run the code?

Prepare data

• Follow the instructions in meijieru/crnn.pytorch to create lmdb datasets. Use the same step to create train and val data.

Change parameters and alphabets

Please update the parameters and alphabets according to the requirement.

• Change parameters in the mytrain.py file

• Change alphabets

  Please put all the alphabets that appear in your labels in a file and input the list as charlist to mytrain.py, else the program will throw an error during training.

Train

Run mytrain.py -

python3 mytrain.py --trainRoot /ssd_scratch/cvit/sanjana/hindi-train-lmdb \
--valRoot /ssd_scratch/cvit/sanjana/hindi-test-lmdb \
--arch crnn --lan hindi --charlist /ssd_scratch/cvit/sanjana/crnn_new/lexicon.txt \
--batchSize 32 --nepoch 15 --cuda --expr_dir /ssd_scratch/cvit/sanjana \
--displayInterval 10 --valInterval 100 --adadelta \ 
--manualSeed 1234 --random_sample --deal_with_lossnan 

Reference

meijieru/crnn.pytorch
Sierkinhane/crnn_chinese_characters_rec

If you use the dataset or code from this work, please add the following citation:-

@inproceedings{gunnaNonLatin2021,
  title={Towards {B}oosting the {A}ccuracy of {N}on-{L}atin {S}cene {T}ext {R}ecognition,
  author={Sanjana Gunna and Rohit Saluja and C V Jawahar},
  booktitle={2021 International Conference on Document Analysis and Recognition Workshops (ICDARW)},
  year={2021},
  organization={IEEE}
}