DevanagariOCR

An End-End DL based model on Devanagari OCR

The pipeline consists two phases: Text Detection and Text Recognition. For Text detection, EAST model trained on ICDAR 2013 and ICDAR 2015 dataset is used which is later fine tuned for devanagari language by changing the threshold values. For Text Recognition, CRNN model is used which is trained on IIITH Hindi dataset

The implementation of EAST: Efficient and Accurate Scene Text Detector is inspired from this repository.

Requirements

Keras 2.0 or higher, and TensorFlow 1.0 or higher should be enough

Data

The trained EAST model can be downloaded from here. The trained CRNN model can be downloaded from here.

Test

Both the models should be in models/ directory. The document images that you want to recognize should be in Input/ directory. One can change the path of the directories in the arguments in OCR.py. The output of the model consists image with bounding boxes drawn on it, bounding boxes coordinates and predicted text word document in the Output/ directory. For running script, type the following command in the terminal or command prompt.

python OCR.py

Results of CRNN Model on IIITH Dataset

Character level Accuracy: 92.44%

Word level Accuracy: 71.56%

Lexicon based Character level Accuracy: 95.89%

Lexicon based Word level Accuracy: 90.02%