This is a number recognition software made from scrath in cpp. No external libraries are used apart from opencv which is only used as a way to visually interact with the program.
If you wish to use the gui, before compliling be sure to include the opencv library.
make all- Binaries are stored in bin/
- What each binary does should be clear from the name
- When train is done running, data from the network in saved in data/
- Functions are provided to save, load and interact with the network
The program uses a neural network to predict a number drawn on an one channel 28x28 image in which pixels range from 0 - 255. A trained neural network takes the image as an input and outputs out what number it thinks it is between 0 - 9. Training the network is done through a process called backpropagation, which is just gradient descent on steroids.
Even though the program is 95% accurate with training and 92% accurate with testing data, this approach is not ideal. The network is very limited on what data can be inputted and what it can be used for. The network is poor at predicting numbers drawn outside of the center of the 28x28 grid. To amend this, we can change the progam to run an input image through a CNN(Convolutional Neural Network) and predict the number based on the structure of the input rather than the pixel values.
- (All the math you need to know) https://explained.ai/matrix-calculus/
- (opencv for linux) https://docs.opencv.org/4.x/d7/d9f/tutorial_linux_install.html
- (opencv for windows) https://opencv.org/releases/
- (MNIST database) http://yann.lecun.com/exdb/mnist/
- https://www.youtube.com/watch?v=aircAruvnKk&t=1037s
