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

Table 8

Comparison of estimation accuracy for crack-detection methods.

Image-set	DNN methods	Metric (%)			FPS
		ODS	OIS	AP
Underground tunnel test set (total 20 images)	MSML mask DCNN w/four scale layers	34.22	35.41	11.88	7.05
	MSML mask DCNN w/four scale layers w/LSI	44.31	42.94	20.21	5.98
	Surf & CNN-based classification model [13]	9.12	7.18	4.59	19.65
	Mask R-CNN [14]	10.91	9.69	7.53	23.91
	Dual-scale CNN-based classification [15] (GoogLeNet & ResNet)	22.81	19.32	10.24	10.98
	CrackPix [16]	25.18	29.33	9.57	14.69

(Best scores shown in bold font).