In scikit-learn SVC, there are 4 types of kernels as shown below. The non-linear kernels do dot product using the kernel functions. 

HSV stands for Hue, Saturation, Value, which is another representation of an image. We get 3 additional histograms from HSV images, by resizing to 16*16. 

• Take x and y direction gradient. 

• Resize to 16*16. 

• Apply Canny Edge Detector to the image. 

• Resize to 16*16. 

Using hog function from scikit-learn. HOG counts occurance of gradients orientation in the localized portion of an image. It contains both angle and magnitude information. 

As you may or may not have seen. NST style transformation done by our project uses SqueezeNet. Some modules in SqueezeNet contain important information of images. So, here we computed 4 gram matrices for each image and treat them as features. Since gram matrices are symmetrical, so we only need half of it. 

After many trials (almost 15 versions), the best combination of classification so far is to have total feature list length of ~1700, with the following features: 

• RGB Histogram

• Grayscale Histogram

• HSV Histogram

• Sharpness Histogram

• Edge Length Histogram

• Histogram of Oriented Gradients

• SqueezeNet features

All the results below use SVC with rbf kernel and C=1. 

• Accuracy: 48.36%

• Training time: 16 minutes

As we can see above, "Expressionism", "Post-Impressionism", and "Realism" are difficult for current SVM to classify. If we remove those hard ones, we have:

• Accuracy:  63%

• Training time: 9 minutes

If we classify only 2 categories at a time, we have pretty high accuracies: