# Multi-scale and Context-adaptive Entropy Model for Image Compression

arXiv: https://arxiv.org/abs/1910.07844 · HTML: https://mlsanity.com/paper/1910.07844/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 17 Oct 2019.

## Built on by (most cited first, 9 of 9)

- [Checkerboard Context Model for Efficient Learned Image Compression](https://mlsanity.com/paper/2103.15306/) (2021)
- [Contextformer: A Transformer with Spatio-Channel Attention for Context Modeling in Learned Image Compression](https://mlsanity.com/paper/2203.02452/) (2022)
- [Selective compression learning of latent representations for variable-rate image compression](https://mlsanity.com/paper/2211.04104/) (2022)
- [Attention-based Convolutional Autoencoders for 3D-Variational Data Assimilation](https://mlsanity.com/paper/2101.02121/) (2021)
- [Efficient Contextformer: Spatio-Channel Window Attention for Fast Context Modeling in Learned Image Compression](https://mlsanity.com/paper/2306.14287/) (2023)
- [Multi-Context Dual Hyper-Prior Neural Image Compression](https://mlsanity.com/paper/2309.10799/) (2023)
- [DeepHQ: Learned Hierarchical Quantizer for Progressive Deep Image Coding](https://mlsanity.com/paper/2408.12150/) (2024)
- [Exploring the Rate-Distortion-Complexity Optimization in Neural Image Compression](https://mlsanity.com/paper/2305.07678/) (2023)
- [Approximately Invertible Neural Network for Learned Image Compression](https://mlsanity.com/paper/2408.17073/) (2024)

## Abstract (arXiv)

We propose an end-to-end trainable image compression framework with a multi-scale and context-adaptive entropy model, especially for low bitrate compression. Due to the success of autoregressive priors in probabilistic generative model, the complementary combination of autoregressive and hierarchical priors can estimate the distribution of each latent representation accurately. Based on this combination, we firstly propose a multi-scale masked convolutional network as our autoregressive model. Secondly, for the significant computational penalty of generative model, we focus on decoded representations covered by receptive field, and skip full zero latents in arithmetic codec. At last, according to the low-rate compression's constraint in CLIC-2019, we use a method to maximize MS-SSIM by allocating bitrate for each image.