# YUV Paper

### 1. [Transform Network Architectures for Deep Learning based End-to-End Image/Video Coding in Subsampled Color paces](https://arxiv.org/abs/2103.01760)
* IEEE Open Journal of Signal Processing

**1.** based on input-output channel alignment
* Which aim to support YUV420 without introducing any major changes to the existing network architectures.
* Separate-channel Coding
>Y, U and V components are split channel-wise, using single and double encoder-input/decoder-output channel.
>
* Six-channel Coding
>Luma component(Y) is spatially split into four channels by matching the dimensions of chroma components. This approach only requires training a single network.
>

**2.** Proposed Transform Network Architecture
* Branched network structure
> chroma branch does not apply any downsampling
* Cross-channel transformation
>1×1 convolutional layer is used to combine the output of branched network carrying luma and chroma information separately at the encoder side.
* Activation functions
>Change activation into PReLU.

### 2. [Learned Variable-Rate Multi-Frequency Image Compression using Modulated Generalized Octave Convolution](https://arxiv.org/abs/2009.13074)
* MMSP 2020
* In terms of YUV PSNR, the scheme is very similar to HEVC.
* Generalized from the recently developed octave convolution, where feature maps are factorized into high-frequency (HF) and low-frequency (LF) components, and the LF components have lower resolution than the HF components.