Browse Teach Curate

Entropy-Gradient Proxy and Optimal Vector Quantization for Neural Image Codecs

Two inference-time techniques for improving off-the-shelf pretrained neural image and video codecs without retraining. A predefined optimal uniform vector quantization replaces scalar quantization of the latent, and the entropy gradient available at the decoder is used as a proxy for the reconstruction-error gradient, which the decoder cannot compute, to refine the latent and lower the bitrate at equal quality.

Rosie Text

IEEE TRANSACTIONS ON IMAGE PROCESSING 1 Exploiting Latent Properties to Optimize Neural Codecs

End-to-end neural image and video codecs are rate-distortion autoencoders that jointly train an encoder, a decoder, and an entropy model, so that a latent representation is quantized and entropy-code…