VietnameseOCR - Vietnamese Optical Character Recognition

Apply Deep Learning ( CNN networks ) to train a model uses for recognizing Vietnamese characters, it works well with Latin characters.

Dataset in big image ( 10.000 samples, 2800 x 2800 pixel)

Requirements

python 3.6.5
tensorflow
PIL

Model Summary

| Layer | Shape | Kernel | Stride | Padding | | | -------------:| --------------:|---------------:|--------------:|------------:|-----------:| | INPUT | [28, 28, 1] | | | | | | CONV1 | | [3, 3, 32, 32] | [1, 1] | SAME | | | POOL1 | | | | | | | CONV2 | | [3, 3, 32, 64] | [1, 1] | SAME | | | POOL2 | | | | | | | CONV3 | | [3, 3, 64, 128]| [1, 1] | SAME | | | POOL3 | | | | | | | FC1 | | | | | | | FC2 | [625, 190] | | | | |

Results

Training...
...... Epoch: 38 cost = 0.312853018 Epoch: 39 cost = 0.298816641 Epoch: 40 cost = 0.293328794
Evaluation 
 Test Accuracy: 0.974867469544

Training

Prepare dataset for training

git clone https://github.com/miendinh/VietnameseOCR.git
cd VietnameseOCR/data/train/characters
unzip dataset.zip

Let's train.

python train.py

Create you own dataset

Prepare fonts for generating text-image

• You could add more fonts

cd VietnameseOCR/data/train/characters
unzip google.zip
unzip win.zip

Create font list, then save it in fonts.list

source ./list.sh

Generate Text Image Dataset

python generate_data.py

Play with pretrained model

• All pretrained weights of model is save to file vocr.brain
• Let's test with random character in dataset

python predict.py

Further working

• [x] Character classification.
• [ ] Dataset augmentation.
• [ ] Improve accuracy.
• [ ] Text location.
• [ ] Text recognition.
• [ ] Apply NLP for spell checking.

References

• STN-OCR: A single Neural Network for Text Detection and Text Recognition
• Automatic Dataset Augmentation
• VGG16 implementation in TensorFlow
• Vietnamese Dict (VietOCR3)

Author mien.hust [at] gmail [dot] com