[point cloud compression] variable image compression with a scale hyperprior

Variational Image Compression with A Scale Hyperprior

This paper presents a new method of compression ： Using transcendental knowledge . Transcendental yes ” A priori a priori ”.

Intro

This paper gives the edge information （Side information） The definition of ： Side information is an additional bit stream from the encoder to the decoder , The information is modified to the entropy model , This reduces mismatches （additional bits of information sent from the encoder to the decoder, which signal modifications to the entropy model intended to reduce the mismatch）. therefore , This kind of edge information is regarded as a priori of entropy model parameters , And edge information has become a hidden representation “ A priori a priori ” 了 .

Ideas

Background

Transformation based models

Coding of transformations （Transform coding） Now in the depth of learning is popular . Input the vector of the image $x$ You can use a parameterized transformation , become ：

$$y=g_a(x;{\varphi }_g)$$

there $y$ Is a potential feature ;${\varphi }_g$ Yes converter （ Encoder ） Parameters of ; This process is called Parametric Analysis The process . And notice , there $y$ It needs to be quantized before entropy coding （ Quantized to discrete values , So that it can be entropy encoded losslessly ）. It is assumed that the potential characteristics after quantification are $\hat{y}$, Then the transformation used in the reconstruction , bring ：

$$\hat{x}=g_s\left(\hat{y};{\theta }_g\right)$$

among , This process is called Parametric Synthesis The process （ Here, it can also be regarded as a decoder ）.${\theta }_g$ Is the parameter of the decoder .

VAE

Variational self encoder （Variational Autoencoder, VAE） Compare with AE, It maps the input to a distribution （ This distribution is usually Gussian） Not a specific vector , As described in the previous section Transformation based models Medium $y$. stay VAE in , He used “ Inferential model ”（Inference Model） Deduce the potential representation in the probability source of the image （“inferring” the latent representation from the source image）, use “ Generate models ”（Generative model） Generate the probability to get the reconstructed image .

For more details, please refer to [1]. But notice , In this paper , We use $z$ To express super prior information rather than potential distribution . Please distinguish .

Model

Pictured 2 Shown , Potential representations obtained by using prior knowledge $y$（ chart 2 The second graph from the left of ） There are structural dependencies （ Spatial coupling ）, This cannot be captured by the total decomposition of the variational model . therefore , The model will be modeled in a super prior way .

The so-called super a priori is a priori of a priori . therefore , Then a potential representation is established $y$ Potential representation of $z$, To capture this spatial dependency . It is worth mentioning that , there $z$ That is, edge information （$z$ is then quantized, compressed, and transmitted as side information）. Capture potential representations $z$ after , After quantification $\hat{z}$ To estimate $\hat{\sigma }$. This $\hat{\sigma }$ Will be used to reconstruct at the decoder side $\hat{y}$, In order to obtain $\hat{x}$.

Reference

[1] https://libertydream.github.io/2020/07/26/ from Autoencoder To beta-VAE/# VAE： Variational self encoder