Features from local binary pattern histogram?

Question

I'm trying to determine a correlation amongst some texture samples based on their LBP histograms. Most literature I've been able to find on the subject discusses measuring distances between pairs of histograms (such as Euclidean distance), essentially treating each of the N values of the histogram as a separate feature and trying to cluster within N dimensional space.

I would prefer not to treat each value as a separate feature, as I'd like to combine my data with other texture features before my analysis. I'm wondering if there is a non-comparative feature which I could extract from the histograms instead.

Answer 1

Comparing LBP histograms using a dissimilarity measure is indeed a commonly used approach to LBP-based image classification (see this review on the topic).

Alternatively you could extract features from the LBP histograms themselves, as described in this paper on retinal disease screening through LBP-based analysis of fundus images:

Different statistical information is extracted from these histograms to use it as features in the classification stage. Concretely, the calculated statistical values are: mean, standard deviation, median, entropy, skewness, and kurtosis. To sum up, six statistical values are calculated from each LBP and VAR histogram, giving place to 12 features for each radius used. Consequently, the total number of features is equal to 144 (12 features × 4 radius × 3 components).