Recognition of Handwriting Numbers

This project is an AI agent trained using KNN, Neural Network (NN), and Convolutional Neural Network (CNN) models to recognize handwriting numbers. The numbers are drawn in real-time by holding down the left mouse button in the Pygame window. Please read my project descriptions here https://theo-xiao-sg.github.io/handwriting_recognition.html

Running Guide

This project is based on the Python programming language and primarily utilizes standard libraries like TensorFlow, Pillow, scikit-learn, joblib, pgzero, and pygame. Please note that the Keras library is automatically installed together with TensorFlow as well. Keras is a high-level neural networks API.

Environment Setup

Download the requirements.txt and install the required Python libraries. Please note all my 4 projects share the same requirements.txt. If you have done the installation for one project, you can skip it for the other 3 projects

# install all packages using requirements.txt
python -m pip install -r requirements.txt

Training the Model

• If you want to train your model, you can run KNN_model.py, neural_network_model.py, or CNN_model.py in the folder. The model will train itself using 70% of the total 1447 images in the folder images_training. I have selected the best model parameters for KNN and NN models in the codes but you can always try yourselves with a set of parameters. For KNN_model.py, you can amend the line 29. For neural_network_model.py, you can amend the line 31. For CNN_model.py, I haven't done much of hyperparameter tuning. But I am quite happy with the hyperparameter set since the accuracy is around 98%-99%.

• Then, the trained model file saved in a pickle file, either handwriting_knn.pkl, handwriting_knn.pkl, or cnn.pkl, will be generated in the folder.

Try the recognition tool

• To try CNN model, you need to run testing_cnn.py, and then a pygame window will show up. The CNN model is the best in the three models and acchieve 98-99% accuracy in this case. You should definitely try it.

• To try KNN and NN models, you need to run testing.py since the images are reshaped differently from CNN. In the line 8 of testing.py, I have chosen the model which I prefer and it is a Neural Network model. If you like, you can try the other KNN model or any model you trained by amending the line 7 or 8 of testing.py.

• Hold down the left mouse button and draw a number from 0 to 9, and click the key Detect, then you get the recognized number using the AI model you just trained. I hope it is a correct recognition. For me, most of the tests came with correct answers.

• Then, please click the key Clear to clear the canvas before you try the next number.

• Try all the numbers you like.

• I hope you are impressed by this tool by now!

Results Illustration

• I have made an animated image file handwriting_recog_demo.gif and a video file handwriting_recog_demo.mp4 to illustrate the number recognition tool. Both files are saved in the folder results.

• You can go to the folder results and simply click the file handwriting_recog_demo.gif in Github. You can immediately see how powerful this number recognition tool is.

• If you download everything and use windows pc, you can open handwriting_recog_demo.mp4 by Media Player, or most browsers, including Chrome and Edge.

• If you download everything and use a Windows pc, you can open handwriting_recog_demo.gif by Photos, or most browsers, including Chrome and Edge. You should not use Paint since it only opens the 1st image.

• I hope you are impressed by this tool as I was!