In end-to-end image compression, encoder and decoder are jointly trained to minimize the well known rate distortion trade-off, balancing bit rate and quality. However, each quality requires a separate encoder-decoder pair, increasing storage and switching complexity on user devices. This work introduces STanH, a differentiable quantizer based on a parametric sum of hyperbolic tangents. Implemented as a learnable activation layer, STanH can adapt pre-trained fixed-rate models to different bitrates.
METHOD FOR LEARNED IMAGE COMPRESSION AND RELATED AUTOENCODER
Alberto Presta;Attilio Fiandrotti;Marco Grangetto;Enzo Tartaglione
2024-01-01
Abstract
In end-to-end image compression, encoder and decoder are jointly trained to minimize the well known rate distortion trade-off, balancing bit rate and quality. However, each quality requires a separate encoder-decoder pair, increasing storage and switching complexity on user devices. This work introduces STanH, a differentiable quantizer based on a parametric sum of hyperbolic tangents. Implemented as a learnable activation layer, STanH can adapt pre-trained fixed-rate models to different bitrates.
File in questo prodotto:
| File | Dimensione | Formato | |
|---|---|---|---|
|
SVT076_IT_1a bozza.pdf
Accesso riservato
Dimensione
578.82 kB
Formato
Adobe PDF
|
578.82 kB | Adobe PDF | Visualizza/Apri Richiedi una copia |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
