Table 3.

MATLAB implementation and parameter settings of the PFP-LHCINCA model.

Method	Parameter
Image resizing	256 × 256
Image decomposition	Average pooling with four levels using 2 × 2, 4 × 4, 8 × 8, and 16 × 16
Patch division	16 × 16 sized patches
LPQ and HOG feature extraction	341 (256 LPQ and 36 HOG) features are extracted for each patch
Feature merging	The concatenation function is merged
Chi2	The most informative 1000 features are selected
INCA	Range: [100, 1000]; error function: kNN with 10-fold CV. Herein, k is 1, the distance metric is Euclidean, and weight is none
Classifiers	kNN: k = 70, distance: correlation, weight: squared inverse
	LD: discriminant type: linear, gamma: 0
	NB: kernel: normal, support: unbounded
	SVM: kernel function: Gaussian, box constraint: 3, kernel scale: 5.6
	DT: split criterion: deviance, maximum number of splits: 51, surrogate: off
Bayesian optimizer	Acquisition function: expected improvement per second plus, iterations: 100