This repository contains tool which allow to detect region with text and translate it one by one.
Two pretrained neural networks are used. One of them is responsible for detecting places in which text appear and return its coordinates. Structure use for this operation is based on CRAFT architecture.
Second network take detected words and recognize words included inside it. Convolutional Recurrential neural networks (CRNN) are used for this operation.
Under construction
I decided to deploy it on heroku (temporarily solution), but the amount of memory available on this platform is not enough. You can check it on heroku app. I decided to add bootstrap template because whole solution become more intuitive.
To install it locally, you can run from your virtual env
python -m pip install requirements.txtto install it properly on Linux OS you have to install additionaly
apt-get update
apt-get install -y libsm6 libxext6 libxrender-dev
pip install opencv-python
If problems with cv2 imports are still appearing then you should install
pip install opencv-contrib-python
Then you can run
```python
python -m pip install requirements.txt
To run it locally, please activate your environment
> win
venv\Scripts\activate.bat
>linux
source venv\Scripts\activate
and run straight from project origin
python app.py
If everything goes properly, you'll see on localhost:8000, screen just like one below.
I decided to remove argparse, because as I mention earlier, it was less intuitive. Solution is not fast, is more like an toy example which shows how to use Pytorch model on deployment environment.
Version which I use here contain torch-cpu which make preprocessing and detecting slightly slower. I test it on cuda and it was much faster.
If you have more information, drop me a line If you like it, give a star
Draft: Show how does it work on complex .tif example document.