Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Color Learning for Image Compression

Srivatsa Prativadibhayankaram, Thomas Richter, Heiko Sparenberg, Siegfried Fößel

TL;DR

This work tackles color fidelity in learned image compression by separating structure and color information into luminance and chrominance branches within the YUV space. It introduces a dual-branch end-to-end codec with hyperpriors and attention mechanisms, optimizing a rate-distortion objective that includes a color-difference term: $L = R + lambda1*MSE + lambda2*(1 - MS-SSIM) + lambda3*DeltaE00^{12}$. Channel impulse-response analysis demonstrates separation of structure and color in the learned representations, supporting potential cross-component prediction and residual coding. Empirical results on Kodak and related datasets show color-aware learning yields competitive performance against JPEG, VVC, and other learned codecs, with improved color fidelity.

Abstract

Deep learning based image compression has gained a lot of momentum in recent times. To enable a method that is suitable for image compression and subsequently extended to video compression, we propose a novel deep learning model architecture, where the task of image compression is divided into two sub-tasks, learning structural information from luminance channel and color from chrominance channels. The model has two separate branches to process the luminance and chrominance components. The color difference metric CIEDE2000 is employed in the loss function to optimize the model for color fidelity. We demonstrate the benefits of our approach and compare the performance to other codecs. Additionally, the visualization and analysis of latent channel impulse response is performed.

Color Learning for Image Compression

TL;DR

This work tackles color fidelity in learned image compression by separating structure and color information into luminance and chrominance branches within the YUV space. It introduces a dual-branch end-to-end codec with hyperpriors and attention mechanisms, optimizing a rate-distortion objective that includes a color-difference term: . Channel impulse-response analysis demonstrates separation of structure and color in the learned representations, supporting potential cross-component prediction and residual coding. Empirical results on Kodak and related datasets show color-aware learning yields competitive performance against JPEG, VVC, and other learned codecs, with improved color fidelity.

Abstract

Deep learning based image compression has gained a lot of momentum in recent times. To enable a method that is suitable for image compression and subsequently extended to video compression, we propose a novel deep learning model architecture, where the task of image compression is divided into two sub-tasks, learning structural information from luminance channel and color from chrominance channels. The model has two separate branches to process the luminance and chrominance components. The color difference metric CIEDE2000 is employed in the loss function to optimize the model for color fidelity. We demonstrate the benefits of our approach and compare the performance to other codecs. Additionally, the visualization and analysis of latent channel impulse response is performed.
Paper Structure (11 sections, 2 equations, 4 figures)

This paper contains 11 sections, 2 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Deep Neural Network based Color Learning
  3. Model and workflow
  4. Optimization
  5. Channel impulse response
  6. Implementation details
  7. Results
  8. Comparison of R-D curves
  9. Comparison of reconstruction fidelity
  10. Comparison of channel impulse responses
  11. Conclusion

Figures (4)

  • Figure 1: Network architecture of the proposed color learning model. The upper part indicates the luminance branch and the lower part is the chrominance branch.
  • Figure 2: R-D curves of the proposed model (ours), bmshj2018balle2018variational, cheng2020cheng_learned_2020, JPEG 125072 and VVC 9503377 for the Kodak dataset.
  • Figure 3: Comparison of original image patch and decoded patches from the proposed model, JPEG and cheng2020, compressed with a bitrate of around 0.4 bpp.
  • Figure 4: Channel impulse responses for the image ClassA_8bit_Bike_2048x2560_8b_RGB.png. The sub-images of size $16\times16$ are arranged in a grid based on the decreasing order of channel bitrate contribution.