c++ - Feature extraction: Divide the ROI into small windows(patches) or let it be as ROI? -

i interested in feature extraction of roi of image, used feature vectors input svm. features extracted include texture(color co-occurence/haralick features) , color(rgb-histogram, 4 bins , mean r, g, b)- total of 29 features/attributes.

question:
there difference if feature extraction(computation) done directly on roi image compared dividing roi small parts(nxn window/patch) first , computing features? advantages/disadvantages? of 2 give more meaningful features?

let's if dividing roi small parts better option , fed svm in training stage, how svm-predict or testing like? feature extraction of test images/test region of interest have divided small parts , fed svm-predict function(just training process)? or feature extraction on test roi directly do?

by way, i'm trying classify disease type found in leaf(3 types, , 1 healthy). , far, got poor results in svm-predict(predicts 1 class, around 20%) , found out features each class not distinguished each other. here's sample images of roi(3 types).

3 types: https://drive.google.com/folderview?id=0b1axcxzd_oadx3vvbjlkbfjbzzg&usp=sharing

i'm using opencv. in c++. explanation , reference helpful.

features remain same or different depends upon type of features going use. example, if features hue-histogram don't think affect if use features sift/surf surely different because sift consider several factors edges , might possible en edge in whole image won't appear edge in smaller region.

i suggest conduct tests because theoretical results , practical results differ. so, try find how features if extract them using whole roi....and how in total if extract them after dividing roi several parts.