Pixel-Level Recognition of Pavement Distresses Based on U-Net

Table 3

The architecture of U-net based on ResNet-34.

The contraction path	Layers	The expansion path	Layers
Conv2d	[7 × 7, 64], stride = 2	Pixel shuffle	[1 × 1, 1024]
Max pool	[3 × 3], stride = 2	Average pool	[2 × 2]
Conv2d		Conv2d
Conv2d		Pixel shuffle	[1 × 1, 512]
Downsample	[1 × 1, 128]	Average pool	[2 × 2]
Conv2d		Conv2d
Conv2d		Pixel shuffle	[1 × 1, 768]
Downsample	[1 × 1, 256], stride = 2	Average pool	[2 × 2]
Conv2d		Conv2d
Conv2d		Pixel shuffle	[1 × 1, 512]
Downsample	[1 × 1, 512], stride = 2	Average pool	[2 × 2]
Conv2d		Conv2d
Conv2d	[3 × 3, 1024]	Pixel shuffle	[1 × 1, 384]
Conv2d	[3 × 3, 512 ]	Average pool	[2 × 2]
		Conv2d	[3 × 3, 49]
		Conv2d	[3 × 3, 99]
		Conv2d	[1 × 1, 7 ]