An experiment in generating ASCII Art from images with an artificial neural network.
IN PROGRESS
Currently only a hard-coded test image is converted:
$ python test_ocr.py
Source:
Generated:
This experiment trains a neural network model as an optical character recognizer.
The model is a simple 2 layer neural network:
- 99 inputs (glyphs in the font are 9x11 pixels)
- fully connected hidden layer with 99 nodes with bias and sigmoid activation
- fully connected output layer with 92 nodes (number of different characters used) with bias and softmax activation
- cross entrophy loss function
It is trained with rendered glyphs of a font and corresponding labels as inputs, e.g.:
 |
 |
 |
|---|---|---|
| a | b | # |
And uses those to predict the best label for the tiles in an input image, e.g.:
 |
 |
 |
 |
 |
|---|---|---|---|---|
| _ | , | # | * | 7 |
Two implementations of the model were made:
This is my own implementation of MLPs using only numpy arrays. It supports batch learning with stochastic gradient descent.
Developing this has been a great help in understanding the maths behind the back propagation algorithm and why it is so efficient for calculating the loss derivates necessary for gradient descent. I highly recommend the blog post Calculus on Computational Graphs: Backpropagation for further reading.
This is an implementation using the Keras Deep Learning Library. It is trained using stochastic gradient descent with Nesterov momentum.
Training this model is quite a bit faster then for my numpy implementation and really simple to implement.
- Use deep convolutional neural nets to detect more abstract features for better selection of ASCII character
- generate more general training data by randomly transforming the glyphs (translation, scale, shear)
- use overlapping segments of input image to include surrounding pixels into selection of ASCII character