Study on the Method of Fundus Image Generation Based on Improved GAN

Table 1

Network structure of the generator G₂.

	Layer	Input	Output
Downsample	Conv (64, 4 × 4), BN, LReLU	(128, 128, 1)	(128, 128, 64)
	Conv (128, 4 × 4), BN, LReLU	(128, 128, 64)	(64, 64, 128)
	Conv (256, 4 × 4), BN, LReLU	(64, 64, 128)	(32, 32, 256)
Residual block	Conv (256, 4 × 4), BN, LReLU	(32, 32, 256)	(32, 32, 256)
	Conv (256, 4 × 4), BN, LReLU	(32, 32, 256)	(32, 32, 256)
	Conv (256, 4 × 4), BN, LReLU	(32, 32, 256)	(32, 32, 256)
	Conv (256, 4 × 4), BN, LReLU	(32, 32, 256)	(32, 32, 256)
Upsample	Deconv (128, 4 × 4), BN, LReLU	(32, 32, 256)	(64, 64, 128)
	Deconv (64, 4 × 4), BN, LReLU	(64, 64, 128)	(128, 128, 64)
	Deconv (32, 4 × 4), BN, LReLU	(128, 128, 64)	(256, 256, 32)
	Conv (1, 4 × 4), BN, LReLU	(256, 256, 32)	(256, 256, 1)