跳转至

DVC: An End-to-end Deep Video Compression Framework

Key Points

end-to-end video compression deep model

learning based optical flow estimation

auto-encoder style neural networks

Rate-distoration optimization(RDO)

context models to learn the adaptive / non-adaptive arithmetic coding method

generalized divisive normalization (GDN): 一种更适合图像重建的归一化层

Proposed Method

Untitled

Step 1. Motion estimation and compression

CNN model to estimate the optical flow

MV encoder-decoder network: an auto-encoder style network to compress the optical flow

Untitled

Step 2. Motion compensation

Untitled

Step 3-4. Transform, quantization and inverse transform

Transform and inverse: highly non-linear residual encoder-decoder network

Quantization: in order to build ab end-to-end training scheme 量化操作不可差分

replace the quantization operation by adding uniform noise in the traning stage

Step 5. Entropy coding

CNNs Bit rate estimation net to obtain the probability distribution of each symbol

通过CNN估计高斯分布的参数

The correct measure for bitrate is the entropy of the corresponding latent representation symbols.

Step 6. Frame reconstruction