Underwater Imaging Formation Model-Embedded Multiscale Deep Neural Network for Underwater Image Enhancement

Table 1

Details of the multiscale network architecture.

Block	Input size	Operation	Configuration
1	128 × 128 × 3	Convolution	K3S1P1C32
	128 × 128 × 32	ReLU	—
2	128 × 128 × 32	Convolution	K3S1P1C32
	128 × 128 × 32	ReLU	—
	128 × 128 × 32	Maxpool	K3S1P1
3	128 × 128 × 64	Convolution	K3S1P1C32
	128 × 128 × 32	ReLU	—
	128 × 128 × 32	Maxpool	K5S1P1
4	128 × 128 × 96	Convolution	K3S1P1C32
	128 × 128 × 32	ReLU	—
	128 × 128 × 32	Maxpool	K7S1P1
5	128 × 128 × 128	Convolution	K3S1P1C32
	128 × 128 × 3	ReLU	—
6	128 × 128 × 9	Convolution	K3S1P1C32
	128 × 128 × 3	ReLU	—

Note. K, kernel size; S, stride; P, padding; C, channel.