search

dkurman / handwriting_recognition

Recognize addresses written on envelops
0 stars 1 forks source link

Handwritten text recognition (CRNN) #18

Open dkurman opened 5 years ago

dkurman commented 5 years ago

30.03.2019 DONE 1. Load the dataset DONE 2. Install python dependencies

  1. Augment the dataset
  2. Adapt the image_ocr.py code
  3. GPU server access
  4. Train the model
  5. Test the model
dkurman commented 5 years ago

python packages (include to the requirements.txt): numpy==1.16.2 matplotlib==3.0.3 opencv-python==3.4.5.20 tensorflow==1.13.1 jupyter imgaug==0.2.8

bosskairat commented 5 years ago

SimpleHTR (https://github.com/githubharald/SimpleHTR) model was trained. Results: Model 1 with augmentation (bestpath): Character error rate: 17.979%. Word accuracy: 57.111% with augmentation (beamsearch): Character error rate: 17.731%. Word accuracy: 58.333% with augmentation (wordbeamsearch): Character error rate: 15.788%. Word accuracy: 75.111% Model 2 with augmentation and deslanting (bestpath): Character error rate: 19.134%. Word accuracy: 52.555% with augmentation and deslanting (beamsearch): Character error rate: 18.994%. Word accuracy: 53.333% with augmentation and deslanting (wordbeamsearch): Character error rate: 16.382%. Word accuracy: 73.555%.