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

Table 4

Comparison of estimation accuracy for deep convolutional neural networks.

Image-set	DNN method	Metric (%)			FPS
		ODS	OIS	AP
CrackTree260	Mask R-CNN	87.93	93.85	82.81	24.73
	ResUNet	97.01	97.07	94.63	21.11
	MS mask DCNN w/three scale layers	90.53	90.75	90.95	30.31
	MS mask DCNN w/four scale layers	91.59	91.91	90.16	15.27
	MSML mask DCNN w/three scale layers	91.87	91.75	90.02	21.32
	MSML mask DCNN w/four scale layers	92.66	92.98	91.14	7.05
CRKWH100	Mask R-CNN	89.32	89.43	88.73	25.01
	ResUNet	96.41	96.63	94.24	19.21
	MS mask DCNN w/three scale layers	93.96	94.58	97.59	29.12
	MS mask DCNN w/four scale layers	95.64	96.13	97.66	14.56
	MSML mask DCNN w/three scale layers	96.27	96.31	96.89	20.37
	MSML mask DCNN w/four scale layers	96.54	97.03	98.55	6.91
CrackLS315	Mask R-CNN	84.68	85.98	79.65	25.55
	ResUNet	90.31	87.18	83.39	22.37
	MS mask DCNN w/three scale layers	90.05	88.98	90.39	30.33
	MS mask DCNN w/four scale layers	90.03	88.33	89.75	14.78
	MSML mask DCNN w/three scale layers	90.72	87.73	89.01	21.10
	MSML mask DCNN w/four scale layers	91.60	89.17	92.88	7.02
Stone331	Mask R-CNN	58.73	59.37	47.35	24.21
	ResUNet	83.21	80.11	72.65	22.13
	MS mask DCNN w/three scale layers	81.32	79.54	78.03	30.98
	MS mask DCNN w/four scale layers	82.03	81.38	78.82	14.97
	MSML mask DCNN w/three scale layers	83.54	83.89	80.35	21.34
	MSML mask DCNN w/four scale layers	84.71	84.40	83.26	7.05

(Best scores shown in bold font).