Accès libre

Improved Double Regression Nonlinear Image Super Resolution Model

Jieyi Lv

Jieyi Lv

Xi'an Technological University, State and Provincial Joint Engineering Lab. of Advanced Network, Monitoring and ControlXi'an, China
Recherchez cet auteur sur
Sciendo|Google Scholar
Lv, Jieyi
 et 
Zhongsheng Wang

Zhongsheng Wang

Xi'an Technological University, State and Provincial Joint Engineering Lab. of Advanced Network, Monitoring and ControlXi'an, China
Recherchez cet auteur sur
Sciendo|Google Scholar
Wang, Zhongsheng
  
16 août 2023
Improved Double Regression Nonlinear Image Super Resolution Model's Cover Image
À propos de cet article
Article précédent
Article suivant
Citez
Télécharger la couverture
Publié en ligne: 16 août 2023
Pages: 46 - 53
DOI: https://doi.org/10.2478/ijanmc-2023-0055
Mots clés
© 2023 Jieyi Lv et al., published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Figure 1.

U-Net Network structure
U-Net Network structure

Figure 2.

Residual learning unit RB
Residual learning unit RB

Figure 3.

Channel attention (CA)
Channel attention (CA)

Figure 4.

Residual channel attention block (RCAB)
Residual channel attention block (RCAB)

Figure 5.

Double regression theoretical model
Double regression theoretical model

Figure 6.

A double regression network training model based on U-Net network transformation is presented
A double regression network training model based on U-Net network transformation is presented

Figure 7.

Sample graph of data set
Sample graph of data set

Figure 8.

PSNR data comparison graph
PSNR data comparison graph

Figure 9.

Comparison of ppt details
Comparison of ppt details

Figure 10.

Comparison of baby details
Comparison of baby details

Ablation experiment

PSNR CA RB Direct connection path PReLu
37.850 No Yes Yes Yes
37.844 Yes No Yes Yes
37.738 No No Yes Yes
37.919 Yes Yes No Yes
37.941 Yes Yes Yes No
37.978 Yes Yes Yes Yes

Adaptation Algorithm on Unpaired Data

Input: Unpaired real-world data: Su;
  Paired synthetic data: Sp;
  Batch sizes for Su and Sp : x and y;
  Indicator function: Sm.
  1. Load the pretrained models P and u;
  2. while not convergent do
  3. Sample unlabeled data {xi} from SU;
  4. Sample labeled data {(xi, yi)} from SP;
  5. // Update the primal mode
  6. Update P by minimizing the objective:
  7. ∑i=1m+nIsp(xi)ιp(P(xi),yi)+λιD(D(P(xi)),xi) \sum\limits_{i = 1}^{m + n} {{I_{{s_p}}}\left( {{x_i}} \right){\iota _p}\left( {P\left( {{x_i}} \right),{y_i}} \right) + \lambda {\iota _D}\left( {D\left( {P\left( {{x_i}} \right)} \right),{x_i}} \right)}
  8. // Update the dual model
  9. Update D by minimizing the objective:
  10. ∑i=1m+nλιp(D(P(xi)),xi) \sum\limits_{i = 1}^{m + n} {\lambda {\iota _p}\left( {D\left( {P\left( {{x_i}} \right)} \right),{x_i}} \right)}
  11. END

Comparison of algorithms for different data sets

Method Set5PSNR/SSIM Set14PSNR/SSIM BSD100PSNR/SSIM
Bicubic 32.40/0.9589 31.32/0.9521 32.87/0.9563
SRCNN 33.36/0.9460 33.78/0.9366 33.57/0.9423
DRCN 33.57/0.9432 33.99/0.9419 33.66/0.9410
ESPCN 34.12/0.9439 34.26/0.9412 34.23/0.9356
SRGAN 34.26/0.9356 34.89/0.9256 34.56/0.9246
Ours 34.12/0.9326 34.56/0.9247 34.57/0.9232
Langue:
Anglais
Périodicité:
4 fois par an
Sujets de la revue:
Informatique, Informatique, autres
RSS Feed de la revue