Handwriting-Recognition

try to learn and get good results for HWR

Model

• Using the EMNIST letter split with augmentation I got a good Test Accuracy: 93.35% but on separate letters
• you can see the training code in the Notebook
• then saved the model as ONNX for speeding up the GUI and make the application lightweight

Word Recognition

• I took the approach of detecting the boundaries of each letter in the image, making a batch to model, getting predictions, and combining the output (for more flexibility).
• One problem for letters that contain multiple parts like i I merge the 2 bbox (if there is overlap horizontally and close enough vertically).

GUI

• I made a simple GUI that has a canvas to draw on the word then predict the word and show the bboxes of the letters

Installing

• first clone the repo git clone https://github.com/Abdo-Eid/Handwriting-Recognition.git
• then change the directory cd Handwriting-Recognition
• then install the requirements pip install -r requirements.txt