The Performance Research of the Data Augmentation Method for Image Classification

Table 1

The network architecture of AlexNet, VGG16, and InceptionV3 used in the experiments. Inception blocks A, B, C,D, and E are shown in the appendix.

AlexNet		VGG16		InceptionV3
Stage	Patch size/stride/or remarks	Stage	Patch size/stride/or remarks	Stage	Patch size/stride/or remarks
Conv		Conv	2 times 3 × 3 × 64	Conv	3 × 3 × 32/2
Max pool		Max pool	2 × 2/2	Conv	3 × 3 × 32/1
Normalization		Conv	2 times 3 × 3 × 128	Conv padded	3 × 3 × 64/1
Conv padded	5 × 5 × 256/2	Max pool	2 × 2/2	Max pool	3 × 3/2
Max pool	3 × 3/2	Conv	2 times 3 × 3 × 256	Conv	1 × 1 × 80/1
Normalization		Max pool	2 × 2/2	Conv	3 × 3 × 192/1
Conv padded	3 × 3 × 384/1	Conv	3 × 3 × 512	InceptionA	3 times; Channel 32/64/64
Conv padded	3 × 3 × 384/1	Conv	2 times 3 × 3 × 512	InceptionB	1 times
Conv padded	3 × 3 × 256/1	Max pool	2 × 2/2	InceptionC	4 time; Channel 128/160/160/192
Max pool	3 × 3/2	Conv	3times 3 × 3 × 512	InceptionD	1
FC	4096	Max pool		InceptionE	2
FC	4096	FC	4096	Global average pool
FC	1000	FC	4096	FC	1000
Softmax		FC	1000	Softmax
		Softmax