False positives reveal relationships among impressionist painters

Abstract

We test if we can classify the works of five (5) French impressionist painters using Convolutional Neural Network (CNN) features derived from different color channels as inputs to a random forest classifier. CNN features is often limited to the gray channel of the image, here we explored features from other color channels. By feeding concatenated features from different color channels, random forest was able to separate the works of these artists to a precision of 63% and recall of 57%. Possible reason to the relatively low metric scores can be attributed to the claim that artists tend to copy the style or draw inspiration from the works of another artists. Interestingly, the false positives from our confusion matrix reveal the similarities between the styles of the painters which agree with the historical relationships between these artists.