Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Traditional Transformation Theory Guided Model for Learned Image Compression

Zhiyuan Li, Chenyang Ge, Shun Li

TL;DR

This paper targets ultra low bitrate image compression and presents an ultra low bitrate enhanced invertible encoding network guided by traditional transformation theory. By integrating Block Discrete Cosine Transformation for feature sparsity and Haar transformation for robust reconstruction within an invertible neural network framework and a hyperprior-based entropy model, it achieves favorable rate-distortion performance without increasing the bitstream. The key contributions are the BDCT layer for sparsity modeling, the Haar downsampling to reduce block artifacts, and the seamless combination with an invertible architecture to minimize information loss. Experimental results on a self-built dataset and Kodak demonstrate superior performance at ultra low bitrates, with clear qualitative improvements in texture and edge preservation over state-of-the-art learned and traditional methods.

Abstract

Recently, many deep image compression methods have been proposed and achieved remarkable performance. However, these methods are dedicated to optimizing the compression performance and speed at medium and high bitrates, while research on ultra low bitrates is limited. In this work, we propose a ultra low bitrates enhanced invertible encoding network guided by traditional transformation theory, experiments show that our codec outperforms existing methods in both compression and reconstruction performance. Specifically, we introduce the Block Discrete Cosine Transformation to model the sparsity of features and employ traditional Haar transformation to improve the reconstruction performance of the model without increasing the bitstream cost.

Traditional Transformation Theory Guided Model for Learned Image Compression

TL;DR

This paper targets ultra low bitrate image compression and presents an ultra low bitrate enhanced invertible encoding network guided by traditional transformation theory. By integrating Block Discrete Cosine Transformation for feature sparsity and Haar transformation for robust reconstruction within an invertible neural network framework and a hyperprior-based entropy model, it achieves favorable rate-distortion performance without increasing the bitstream. The key contributions are the BDCT layer for sparsity modeling, the Haar downsampling to reduce block artifacts, and the seamless combination with an invertible architecture to minimize information loss. Experimental results on a self-built dataset and Kodak demonstrate superior performance at ultra low bitrates, with clear qualitative improvements in texture and edge preservation over state-of-the-art learned and traditional methods.

Abstract

Recently, many deep image compression methods have been proposed and achieved remarkable performance. However, these methods are dedicated to optimizing the compression performance and speed at medium and high bitrates, while research on ultra low bitrates is limited. In this work, we propose a ultra low bitrates enhanced invertible encoding network guided by traditional transformation theory, experiments show that our codec outperforms existing methods in both compression and reconstruction performance. Specifically, we introduce the Block Discrete Cosine Transformation to model the sparsity of features and employ traditional Haar transformation to improve the reconstruction performance of the model without increasing the bitstream cost.
Paper Structure (24 sections, 3 equations, 8 figures, 1 table)

This paper contains 24 sections, 3 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Traditional Image Compression
  4. Learned Image Compression
  5. Methods
  6. Overview
  7. INN Architecture
  8. Haar Transformation
  9. Coupling Layer
  10. Block Discrete Cosine Transformation
  11. Entropy Model
  12. Experiments
  13. Experimental Setup
  14. Datasets
  15. Training details
  16. ...and 9 more sections

Figures (8)

  • Figure 1: Overview of the proposed image compression method
  • Figure 2: Illustration for Block Discrete Cosine Transformation and its inverse process. (a) is the process of BDCT, and (b) is the inverse process of BDCT
  • Figure 3: Example images in our self-built dataset
  • Figure 4: Performance evaluation on the self-built dataset
  • Figure 5: Performance evaluation on the Kodak dataset
  • ...and 3 more figures