Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Perceptual-oriented Learned Image Compression with Dynamic Kernel

Nianxiang Fu, Junxi Zhang, Huairui Wang, Zhenzhong Chen

TL;DR

PO-DKIC addresses the need for perceptually faithful image compression under bitrate constraints. It builds on DKIC by introducing a dynamic residual block group with content-adaptive kernels and an asymmetric space-channel entropy model, plus a GAN-based perceptual loss via PatchGAN and LPIPS. It formulates a constrained optimization using multiple lambda-trained models and linear integer programming to allocate bitrate for maximum perceptual quality. Experiments on standard benchmarks show superior perceptual quality and texture fidelity compared with state-of-the-art methods.

Abstract

In this paper, we extend our prior research named DKIC and propose the perceptual-oriented learned image compression method, PO-DKIC. Specifically, DKIC adopts a dynamic kernel-based dynamic residual block group to enhance the transform coding and an asymmetric space-channel context entropy model to facilitate the estimation of gaussian parameters. Based on DKIC, PO-DKIC introduces PatchGAN and LPIPS loss to enhance visual quality. Furthermore, to maximize the overall perceptual quality under a rate constraint, we formulate this challenge into a constrained programming problem and use the Linear Integer Programming method for resolution. The experiments demonstrate that our proposed method can generate realistic images with richer textures and finer details when compared to state-of-the-art image compression techniques.

Perceptual-oriented Learned Image Compression with Dynamic Kernel

TL;DR

PO-DKIC addresses the need for perceptually faithful image compression under bitrate constraints. It builds on DKIC by introducing a dynamic residual block group with content-adaptive kernels and an asymmetric space-channel entropy model, plus a GAN-based perceptual loss via PatchGAN and LPIPS. It formulates a constrained optimization using multiple lambda-trained models and linear integer programming to allocate bitrate for maximum perceptual quality. Experiments on standard benchmarks show superior perceptual quality and texture fidelity compared with state-of-the-art methods.

Abstract

In this paper, we extend our prior research named DKIC and propose the perceptual-oriented learned image compression method, PO-DKIC. Specifically, DKIC adopts a dynamic kernel-based dynamic residual block group to enhance the transform coding and an asymmetric space-channel context entropy model to facilitate the estimation of gaussian parameters. Based on DKIC, PO-DKIC introduces PatchGAN and LPIPS loss to enhance visual quality. Furthermore, to maximize the overall perceptual quality under a rate constraint, we formulate this challenge into a constrained programming problem and use the Linear Integer Programming method for resolution. The experiments demonstrate that our proposed method can generate realistic images with richer textures and finer details when compared to state-of-the-art image compression techniques.
Paper Structure (10 sections, 6 equations, 3 figures, 1 table)

This paper contains 10 sections, 6 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Architecture
  4. Perceptual Loss Function
  5. Resource Allocation
  6. Experiments
  7. Training Settings
  8. Quantitative Results
  9. Qualitative Results
  10. Conclusions

Figures (3)

  • Figure 1: The framework of our image compression. The upper part is DKIC. The lower part shows the adversarial training.
  • Figure 2: The visual comparison of our method with Cheng2020 and BPG. Our method has more details for the texture of buildings. The parentheses contain (BPP / LPIPS).
  • Figure 3: The visual comparison of our method with Cheng2020 and BPG. Our method has more details for the texture of carrots with yellow rectangles and a clearer texture for the linen with green rectangles. The parentheses contain (BPP / LPIPS).