Handwritten Text Recognition with TensorFlow

Small project on developing a Handwritten Text Recognition (HTR) System over a Streamlit Web Application. Using pre-trained models implemented on TensorFlow and trained on the IAM off-line HTR dataset. These models take images of single words or text lines (multiple words) as input and output the recognized text.

The images are obtained by extracting it from a box (canvas) in which the user can draw or uploading a file with handwritten text on the Streamlit app.

How to use

1. Clone this repo

	git clone https://github.com/Josemvg/HandWriting-Recognition

2. Install the required Python Libraries

	pip install requirements.txt

3. Download the pretrained models and extract them on the model folder, under line-word and model-word folders respectively:
   • Model 1 - Trained on Single Word Images
   • Model 2 - Trained on Text Line Images
4. Next, to run the WebApp execute runner.py or the streamlit app (from the app folder) by using:

	streamlit run webapp.py

5. Once you have the app running, you may now select the Model and Decoder or stick to the Default.

   

   Note: For a tutorial on how to configure Word Beam Search Decoder, please visit the SimpleHTR Repository

   You can start using the app either by writing on the canvas (bear in mind that you can change the stroke width of the pen as well):

   

   Or by uploading an image from your own device:

   

6. Finally, press the Infer button to obtain the most accurate candidate according to the Model, and its probability.

References

Models used on this app pertain to the SimpleHTR Repository by Harald Scheidl.

• How to easily do Handwriting Recognition using Deep Learning
• Build a Handwritten Text Recognition System using TensorFlow
• SimpleHTR - Handwritten Text Recognition with TensorFlow