kpentchev/artfinder

artfinder

ML project for recognizing art paintings.

Technologies used:

• python 3
• tensorflow

Install

• brew install python3
• vi ~/.bash_profile ; alias python='python3'; source ~/.bash_profile
• pip3 install tensorflow
• pip3 install Augmentor

Steps

1. Generate training set using distort.py

• input: a folder containing a few source images of the same painting
• output: a folder containing tousands of distorted images of the painting

2. Generate tensorflow input slices using preprocess.py

• input: one folder per painting with tousands of images; label.txt file with one label per line per image
• output: slices for either training or testing; number and names of slaces depend on number of threads used for preprocessing

3. Train model using train_cnn.py

• input: the training slices
• output: a trained model

4. Evaluate model using train_cnn.py

• input: a trained model
• output: accuracy score

Neuronal Network model

A convolutional neuronal network (CNN) with 7 layers:

• 4 convolutional layers
  • each with a filter size of 3, step 1, padding 'SAME'
  • number of filters are 64, 64, 128 and 256 respectively.
  • each is coupled with a pooling layer of size 2, step 2
• 1 flat layer
• 1 fully connected layer
  • 2048 outputs
• 1 output layer
  • outputs = number of labels

Images are processed in batches of 15 and down/up-sized to 384x384 pixels. For each pixel 3 channels (RGB) are used.

Cost function is reduce_mean of softmax_cross_entropy_with_logits_v2. AdamOptimizer with a learning rate of 1e-4 was used for minimizing the function.