I thought this Kaggle challenge would be a good way for me to test myself using what I knew about machine learning and Python's libraries.

I had previously written some simple PNG decoder functions in Python, so I grabbed that (png.py) and used it to:

• Read in the PNG images
• Convert to grayscale
• Highlight green areas to black, and convert the rest of the image to black
• Resize to 32x32

Then I can convert the resized pictures to a feature vector of 1024 (32*32) pixels of either black (0) or white (1).

For the neural network, I used Python's keras. My model consisted of:

• A convolutional layer
• A pooling layer
• A flattening layer
• A dense layer (to output to a vector of length 12, the number of labels)

Fitting the model took about three hours, so I decided I would just run it once and submit my results.

If I were to go back and redo parts of this project, I would try the following:

• Try to trim every image to a single leaf.
• Rotate each image to a standard orientation.

These tweaks would probably increase my accuracy, but since this project was mainly for my own edification, I'm satisfied.

My final model resulted in 67% accuracy on the Kaggle challenge.