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

Table 6

Comparison of estimation accuracy for crack-detection methods.

Image-set	DNN methods	Metric (%)			FPS
		ODS	OIS	AP
Indoor test set (total 10 images)	MSML mask DCNN w/four scale layers	20.12	21.53	5.39	6.98
	MSML mask DCNN w/four scale layers w/LSI	29.86	29.99	12.35	5.10
	Surf & CNN-based classification model [13]	8.55	6.43	1.98	21.58
	Mask R-CNN [14]	9.58	8.59	2.12	23.63
	Dual-scale CNN-based classification [15] (GoogLeNet & ResNet)	15.73	11.28	4.32	10.23
	CrackPix [16]	13.30	10.19	2.40	15.33

(Best scores shown in bold font).