Disclainer

this project does not work. but i'm done trying to figure out where the bug is in this work. if you want an ANN for OCR, there are way better implementations out there. This project was intended for learning purposes only.

Optical Character Recognition via an Artificial Neural Network

Uses one hot encoding to translate a list of 10 arabic numeral images (0-9) and classify them into the cooresponding digit.

Forward propagation uses ReLU for each hidden layer activation function and Softmax for the output layer.

All layers are bipartite.

Files

We are only trying to classify individual numbers.

cmd/translate_dataset/main.go

Image Binarization: Takes the dataset and converts to only be black and white.

cmd/verify_dataset/main.go

Asserts that the images are formatted properly. All images must have the same resolution and only contain 2 colors.

Development

run all tests with:

go test ./...

clean the data so that it only contains 2 colors with:

go run cmd/translate_dataset/main.go

verify the data looks good via:

go run cmd/verify_dataset/main.go

and then run a single pass with:

go run cmd/train/main.go