An Improved Nondestructive Semantic Segmentation Method for Concrete Dam Surface Crack Images with High Resolution

Table 4

Indexes of ONLY-FCN and CNN-FCN by different thresholds on test set.

Method name	Model II threshold	Model B threshold	Model II				Model B			Relative/absolute prediction time per patch
			TP	FP	TN	FN	mR	mP	mIoU
ONLY-FCN	—	0.999	—	29227	—	—	0.188	0.865	0.182	1.00x/7.9 ms
							−76.4%	+33.5%	−66.7%
	—	0.99	—	40719	—	—	0.408	0.806	0.370
							−48.7%	+24.4%	−32.4%
	—	0.9	—	54502	—	—	0.645	0.729	0.514
							−18.9%	+12.5%	−6.0%
	—	0.5	—	69316	—	—	0.795	0.648	0.547
CNN-FCN	0.001	0.5	13496	56779	10364	43	0.795	0.648	0.547	1.42x/11.2 ms
							+0.0%	+0.0%	+0.0%
	0.01	0.5	13410	41299	25844	129	0.795	0.648	0.548	1.22x/9.6 ms
							+0.0%	+0.0%	+0.2%
	0.1	0.5	12885	10988	56155	654	0.798	0.649	0.550	0.84x/6.6 ms
							+0.4%	+0.2%	+0.5%
	0.5	0.5	12151	2851	64292	1388	0.801	0.650	0.552	0.73x/5.8 ms
							+0.8%	+0.3%	+0.9%
CNN	—	—	—	—	—	—	—	—	—	0.53x/4.2 ms

TN denotes the true background predictions.