I am using linear/quadratic discrimination to guess who is the artist of a painting. All the images are RGB format and I use matlab. I have used both original image and images that are resize to be square.
I tried to use features like the average of color channels, average skewness an so on. Then I plot them against each other and so far it is not very easy to use as feature for a linear/qudratic discrimination. Here are examples of the mean and norm of saturation and value channel (hsv in matlab) plotted against eachother for two artist:
As you can see there definitely is a difference in distribution but the error rate is to high.
Do you have any suggestions on what kind of features I can look at that would differentiate the artists even more?
When I say average I mean that I use mean2 to get the average of a channel or tranformation.