Hybrides Implizites Neuronales Netzwerk mit Selektiver Codierung Latenter Variablen

Details

Veröffentlichungs-Nr.

EP4811791

Anmeldetag

18. März 2025

Veröffentlichung

23. September 2026

Rechtsraum

EP

IPC

H04N19/60, H04N19/70

Abstract

Apparatuses and methods are disclosed for encoding and decoding data frames. The encoding comprises training a hybrid INR network to produce a data region of a frame from latent variables. The training determines the parameters of the hybrid INR network based on a coding cost being a function of the latent variables modified by corresponding elements of a mask. The encoding further comprises selecting, using the mask, a subset of the latent variables, and coding into the bitstream the selected subset of the latent variables as well as the parameters of the hybrid INR network and the mask. The decoding comprises decoding from the bitstream the parameters of the hybrid INR network and the mask that indicates the subset of the latent variables. The latent variables are then obtained, where a latent variable belonging to the subset is decoded from the bitstream based on its distribution and a latent variable not belonging to the subset is derived based on its distribution. The data region of the frame is produced by the hybrid INR network from the obtained latent variables using the decoded parameters of the hybrid INR network.

Anmelder

Firma

InterDigital CE Patent Holdings, SAS

Land

🇫🇷 Frankreich

Fachgebiete