Enhancing Structural Crack Detection through a Multiscale Multilevel Mask Deep Convolutional Neural Network and Line Similarity Index

Table 3

Initial range of hyperparameter and optimal value of hyperparameter for each model.

	Hyperparameters
	Early stopping	Adam optimizer
	Patience	Learning rate
	Range of initial settings
	10∼50	10^−5∼10^−2	0.9∼0.999		0.9∼0.999	10^−9∼10^−5
DCNN methods	Optimal values
Mask R-CNN	24	5.67e−4	0.924		0.995	1.81e−7
ResUNet	12	1.27e−3	0.922		0.967	5.12e−8
MS mask DCNN w/three scale layers	11	1.46e−3	0.937		0.977	3.57e−7
MS mask DCNN w/four scale layers	18	8.18e−4	0.920		0.990	1.12e−6
MSML mask DCNN w/three scale layers	25	5.50e−5	0.932		0.975	8.94e−8
MSML mask DCNN w/four scale layers	15	8.21e−4	0.915		0.998	1.12e−7