Improved Crack Detection and Recognition Based on Convolutional Neural Network

Table 1

Network parameter.

Layer	Size	Operator	Filters	Kernel_size	Strides	Padding
Input	227 × 227 × 3	C1	24	20 × 20	2	Same
L1	114 × 114 × 24	P2	—	7 × 7	2	Same
L2	57 × 57 × 24	BN	—	—	—	—
		relu	—	—	—	—
		C3	48	15 × 15	2	Valid
L3	22 × 22 × 48	P4	—	4 × 4	2	Valid
L4	10 × 10 × 48	BN	—	—	—	—
		relu	—	—	—	—
		C4	96	10 × 10	2	Valid
L5	1 × 1 × 96	relu	—	—	—	—
		Drop	—	—	—	—
L6	1 × 1 × 2	C5	2	1 × 1	1	Valid

The learning rate is set to 0.0001, and the learning rate drops to 0.99 for each iteration of 100; the dropout is set to 0.5.