Skip to main content
EURASIP Journal on Wireless Communications and Networking

Table 1 Architectures of the proposed network including three sub-blocks

From: Facial image super-resolution guided by adaptive geometric features

Network architecture

The main SR block

The depth estimation block

The modulation block

Conv: [3×3,3,64]

Conv: [3×3,3,64]

Conv: [3×3,1,64]

Residual: [3×3,64,64]×8

Downsample: [3×3,3,64]×3

 

Feature addition

Residual: [3×3,64,64]×5

 

Feature multiplication

Upsample: [3×3,64,64]

 

Residual: [3×3,64,64]×8

Feature concatenation

 

Feature addition

Upsample: [3×3,64,64]

Conv: [3×3,64,64]×3

Conv: [3×3,64,64]

Feature concatenation

 

Upsample: [3×3,64,64]×2

Upsample: [3×3,64,64]

 

Conv: [3×3,64,3]

Feature concatenation

 

Output image

Conv: [3×3,64,1]

 
  1. Building blocks are shown in brackets (see also Figs. 2 and 3), with the numbers of blocks stacked. Downsampling is performed by convolution with a stride of 2. Upsampling is performed by convolution followed by pixel shuffle. The convolutional layer parameters are denoted as “ <convolutional kernel size >,<input channel size >, and <output channel size >” within square brackets. The ReLU activation function is not shown for brevity
Back to article page