Joint learned-image compression and super resolution

Анотація

Recent advancements in learning-based image coding have demonstrated promising outcomes. These codecs utilize deep neural networks to reduce dimensionality at the stage where a linear transform would typically be applied. This signal representation, known as latent space, can be interpreted by other deep neural networks without decoding, offering benefits for various image processing tasks. In this study, we establish baselines by combining learnedimage compression and Super Resolution (SR) for hybrid modeling. Specifically, we experiment with two baselines: one leveraging fixed image compression and SR models, and the other integrating fixed image compression and adaptively learned-SR models. Experimental results indicate that our second approach yields better perceptual quality than baseline 1. Moreover, our baseline 2 achieved a 23.33% improvement in BD-Rate PSNR and 0.44 dB in BD-PSNR when evaluated on the DIV2K validation set, and a 20.23% gain in BD-Rate PSNR and 0.34 dB in DB-PSNR when evaluated on the JPEGAI test set. Furthermore, we explored the impact of training data sets on the performance of image compression models to determine the optimal choice of training dataset for our hybrid modeling.

Ключові слова: Learned-image compression; image coding; Super Resolution; image enhancement; image resotration

Як цитувати

@article{internauka10816,
  author   = {У, Джуцi},
  title    = {Joint learned-image compression and super resolution},
  journal  = {Міжнародний науковий журнал «Інтернаука»},
  year     = {2025},
  number   = {3},
  doi      = {10.25313/2520-2057-2025-3-10816},
  url      = {https://www.inter-nauka.com/issues/2025/3/10816}
}

TY  - JOUR
AU  - У, Джуцi
TI  - Joint learned-image compression and super resolution
JO  - Міжнародний науковий журнал «Інтернаука»
PY  - 2025
IS  - 3
SN  - 2520-2057
DO  - 10.25313/2520-2057-2025-3-10816
UR  - https://www.inter-nauka.com/issues/2025/3/10816
LA  - en
ER  -