Method for learned image compression and related autoencoder

The method for learned image compression implemented in an autoencoder comprising a learnable encoder (fa) and a decoder (fs), said method comprising the steps of:

• a) extracting from an image (x) a latent space (y) by means of said learnable encoder (fa);
• b) encoding by means of an auxiliary encoder (ha) said latent space (y) to capture spatial dependencies among said latent space (y) to obtain a hyperprior representation (z) of said latent space (y);
• c) quantizing said hyperprior representation (z) by means of a quantizer (U|Q) to obtain a quantized hyperprior representation ({circumflex over (z)});
• d) entropy coding said quantized hyperprior representation ({circumflex over (z)}) by means of an entropy encoder to obtain a bitstream, wherein an entropy model used to encode said quantized hyperprior representation ({circumflex over (z)}) is represented by a probability distribution p{circumflex over (z)};
• e) entropy decoding said bitstream by means of an auxiliary entropy decoder (hs) to obtain an entropy decoded bitstream;
• f) quantizing said entropy decoded bitstream to obtain a quantized latent space (ŷ);
• g) feeding said entropy decoded bitstream to said decoder (fs);
• h) recover a reconstructed image ({circumflex over (x)}) by means of said decoder (fs);
• i) training said autoencoder via standard gradient descent of the backpropagated error gradient by finding learnable parameters (θf,θg) of said learnable encoder (fa) and of said decoder (fs) that minimize a rate distortion cost function L, wherein said entropy encoder is based on a differentiable formulation of a soft frequency counter (SFC).

An autoencoder for learned image compression comprising:

• a learnable encoder (fa) adapted to extract from an image (x) a latent space (y);
• a quantizer (U|Q) adapted to quantize said latent space (y) to obtain a quantized latent space (ŷ);
• an entropy encoder adapted to entropy coding said quantized latent space (ŷ) to obtain a bitstream, wherein an entropy model used to encode said latent space (y) is represented by a probability distribution pŷ;
• an entropy decoder adapted to decode said bitstream;
• a decoder (fs) adapted to recover a reconstructed image ({circumflex over (x)}) from said entropy decoded bitstream;
• said autoencoder being configured to be trained via standard gradient descent of the backpropagated error gradient by finding learnable parameters (θf,θg) of said learnable encoder (fa) and of said decoder (fs) that minimize a rate distortion cost function L, wherein said entropy encoder comprises a soft frequency counter (SFC) based on a differentiable formulation.

An autoencoder for learned image compression comprising:

• a learnable encoder (fa) adapted to extract from an image (x) a latent space (y);
• an auxiliary encoder (ha) adapted to encode said latent space (y) to capture spatial dependencies among said latent space (y) to obtain a hyperprior representation (z) of said latent space (y);
• a quantizer (U|Q) adapted to quantize said hyperprior representation (z) to obtain a quantized hyperprior representation ({circumflex over (z)});
• an entropy encoder adapted to entropy coding said quantized hyperprior representation ({circumflex over (z)}) to obtain a bitstream, wherein an entropy model used to encode said quantized hyperprior representation ({circumflex over (z)}) is represented by a probability distribution p{circumflex over (z)};
• an auxiliary entropy decoder (hs) adapted to entropy decoding said bitstream to obtain an entropy decoded bitstream;
• a second quantizer (U|Q) adapted to quantize said entropy decoded bitstream to obtain a quantized latent space (ŷ);
• a decoder (fs) adapted to recover a reconstructed image ({circumflex over (x)}) from said entropy decoded bitstream;
• said autoencoder being configured to be trained via standard gradient descent of the backpropagated error gradient by finding learnable parameters (θf,θg) of said learnable encoder (fa) and of said decoder (fs) that minimize a rate distortion cost function L, wherein said entropy encoder is based on a differentiable formulation of a soft frequency counter (SFC).