Pixel Food Classificator
Machine Learning Learning Project
Goal of the project: teach neural network Bob to recognize different types of food from drawings without using any ML-related third party libraries.
In Bob's world there are three types of food: burgers, fries and pepsi.
Images of food for training and testing are generated with help of Pillow library in image_generator.py script.
Neural network was trained on datasets consisting of 10-200 32x32 images with 1-10 epochs. Results are on the plots.
Zipped channels
Images was piped into neural network in following format:
data = [R0, G0, B0, ... Ri, Gi, Bi, ... R1024, G1024, B1024]
Grayscale
Here images was converted to grayscale mode
Gi = 0.2126Ri + 0.7152Gi + 0.0722Bi
data = [G0, ... Gi, ... GN]
For some reason grayscale-network does its job better than zip-network when both are trained on small sets. On big sets zip-network classifies with almost 100% probability when grayscale-networks tops out at ~96%.
Usage
Generate images and train NN
python train_network.py
Generate more images and test NN on 'em
python test_network.py
Check options.py to adjust some values that may affect the result