Automated Pavement Crack Damage Detection Using Deep Multiscale Convolutional Features

Table 5

Comparisons of the results with different upsampling features and structures in the upsampling module.

Backbone	Low-level features		3 × 3 Conv structure	OA	mIoU
	Conv1	Block1
ResNet50	✓		(3 × 3, 256) × 1	97.33	70.43
	✓		(3 × 3, 256) × 2	97.55	71.16
	✓		(3 × 3, 256) × 3	97.56	71.28
	✓		(3 × 3,256) × 4	97.60	70.86
		✓	(3 × 3, 256) × 3	97.68	71.68
	✓	✓	(3 × 3, 256) × 3	98.43	71.82
ResNet101	✓	✓	(3 × 3, 256) × 3	98.79	73.53