This project recognizes constellations (dots not stars) drawn in paint (as of now). Haha. The artificial neural network consists of 20 input neurons because the image is 20x20 pixels (the average of every row is the input), 1 hidden layer, 10 hidden neurons (can be configured) and 1 output neuron (I only tried it with one output neuron). This uses forward pass for computing the output and backpropagation for learning. All handcrafted using C#.
Good things about this project
- Can recognize constellations as what I said at the description. But I think this can also recognize other images.
- Backpropagation from scratch.
- Easy to understand backpropagation and can be implemented on other inputs and outputs
Bad things about this project
- Training images is limited to only 2 because it becomes accurate when it is greated than two. I still can't see why, what and debug have I done. I can't understand how to represent the output well. Haha.
Screenshots
