Table 1 Comparison of different waveguide and wave propagation models

From: Synthetic aperture waveguide holography for compact mixed-reality displays with large étendue

Wave propagation model

Waveguide model

PSNR (dB ↑)

Model size (MB)

Training time

 

Number of modes

 

Number of angles

Train

Test

  

Off-axis ASM56

1

–

–

26.6

26.8

–

–

High-order off-axis ASM

1

–

–

27.3

27.4

–

–

High-order off-axis ASM + CNNs34

1

Explicit

1 × 1

31.2

29.6

620

>48 h

High-order off-axis ASM + equations (1) and (2)

1

Explicit

1 × 1

28.9

29.1

48

>24 h

 

1

Explicit

5 × 5

29.8

29.8

1200

>24 h

 

1

Implicit

Continuous

31.7

31.7

11

~12 h

 

3

Implicit

Continuous

33.5

33.3

11

~12 h

 

6

Implicit

Continuous

34.1

33.8

12

~12 h

  1. All models are trained using the same 300 phase–intensity pairs. Peak signal-to-noise ratio (PSNR) is measured for training and test sets. Our implicit network architecture is much smaller in size and requires less training time than existing explicit CNN-based architectures. When used in a low-rank partially coherent setting (that is, with three or six modes), our model achieves significantly better image quality than all coherent models. Bold indicates the best performance—the highest PSNR, smallest model and shortest training time.
Back to article page