Fig. 3.

Network architecture of CAE with multiple hidden layers. Note: CAE with multiple hidden layers corresponds to conventional convolutional network without a pooling layer. In the current study, the number of hidden layers ranged from 1 to 11. The number of feature maps were 20, 30, and 40. To maintain the sizes of the input, the latent representation, and the output at 28 × 28 pixels, the following value was used: kernel, 3 × 3 pixels; padding, 1 × 1 pixels. Abbreviation: CAE, convolutional auto-encoder.