Evaluation of image partitioning strategies for preserving spatial information of cross-sectional micrographs in automated wood recognition of Fagaceae

Journal of Wood Science

Official Journal of the Japan Wood Research Society

Table 3 Comparison of recognition performance

Strategy	Parameters	Dimensionality	F1 score
SIFT	nOctaveLayers = 3, contrastThreshold = 0.06, edgeThreshold = 10, sigma = 1.6 (as default)	128	0.649
BOF	k = 300	300	0.708
SPM	Pyramid level 1	1500	0.722
RSPM	Partition level 2	900	0.738
TSPM	Partition level 0 (=BOF)	300	0.708
RSPM + AC	R2SPM + AC at partition level 7 (k = 18)	1044	0.742
RSPM + PC	R2SPM + PC at partition level 9 (k = 18)	1053	0.746
RSPM + AC + PC	R2SPM + AC and PC at partition level 8 (k = 18)	1215	0.750
VGG16	Optimizer = SGD, learning rate = 0.01	–	0.705

SIFT, scale-invariant feature transform; nOctaveLayers, the number of layers in each octave; contrastThreshold, the contrast threshold used to filter out weak features; edgeThreshold, the threshold used to filter out edge-like features; sigma, the sigma of Gaussian applied to the image at the octave number 0; BOF, bag-of-features; k, number of codewords; SPM, spatial pyramid matching; RSPM, radial-SPM; R2SPM, RSPM with partition level 2; TSPM, tangential-SPM; AC, autocorrelation; PC, Pearson correlation coefficient; SGD, stochastic gradient descent