Deep Learning-Enabled Automatic Detection of Bridges for Promoting Transportation Surveillance under Different Imaging Conditions

Table 1

The mean average precision (mAP) results of four bridge detection models trained with original dataset and enlarged dataset by data augmentation strategy.

Dataset	Methods	Clear	Hazy	Low-light	FPS
Original	Faster R-CNN [9]	55.25 ± 1.17	46.30 ± 1.15	51.11 ± 1.22	9
	YOLOv3 [10]	50.02 ± 0.51	42.24 ± 0.54	47.60 ± 0.47	77
	YOLOv4 [11]	60.05 ± 0.39	51.34 ± 0.42	57.70 ± 0.37	67
	YOLOv5 [12]	60.38 ± 0.35	49.50 ± 0.52	55.55 ± 0.44	140
Data augmentation	Faster R-CNN [9]	57.24 ± 0.33	59.45 ± 0.21	65.09 ± 0.41	8
	YOLOv3 [10]	55.76 ± 0.35	57.34 ± 0.23	63.44 ± 0.28	76
	YOLOv4 [11]	69.38 ± 0.09	68.06 ± 0.11	70.03 ± 0.07	65
	YOLOv5 [12]	60.49 ± 0.14	62.79 ± 0.17	76.47 ± 0.16	142