Patched-Based Swin Transformer Hyperprior for Learned Image Compression
Sibusiso B. Buthelezi, Jules R. Tapamo · 2025
Read Patched-Based Swin Transformer Hyperprior for Learned Image Compression by Sibusiso B. Buthelezi, Jules R. Tapamo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-12842631.txt.
Description
Sibusiso Buthelezi and Jules Tapamo improve learned image compression in this two thousand twenty-five paper. Neural codecs based on autoencoders rely on entropy models, priors over the compressed form, usually built from small networks. Those local priors miss complex global links in the latent space, wasting bits. Full transformer priors capture global structure but cost too much computing for high-resolution images. The authors split the difference with a hybrid design.
Their framework couples a standard autoencoder with a patch-based Swin transformer hyperprior in layers. Shifted window attention models local and global context well, and a smooth quantization module lets the rate-distortion goal train end to end. A better model of the latents yields more compact codes. On the Kodak, JPEG AI, and CLIC datasets, the hybrid beat methods using simpler entropy models on rate-distortion results and visual quality at lower bitrates.
Catalog record
| CREATOR | Sibusiso B. Buthelezi, Jules R. Tapamo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |