Khasi-OCR

Khasi-ocr is a project to create OCR model for khasi language. tesseract-ocr is used for LSTM layer training.

base model: eng.traineddata
output model: kha.traineddata(fast model)
fonts: Liberation Serif
network spec: [1,36,0,1[C3,3Ft16]Mp3,3Lfys64Lfx96Lrx96Lfx192Fc128]
lstmeval result: CER = 0.08, WER = 0.19
UNLV test result: CER = 4.3 (academic textbooks), CER = ~76.5 (dictionary)

Contributors

Uday Kiran Nagineni, Akhilesh Kakolu Ramarao

Improvements

1. edit the groundtruth files manually with reference to images.
2. produce best model of traineddata. use (network spec - Lfx512 O1c1) in lstm training

for more info

refer wiki - khasi-ocr