Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Image Compression with Quantization Rectifier

Wei Luo, Bo Chen

TL;DR

This work tackles the quantization-induced loss of feature expressiveness in neural image compression by introducing a Quantization Rectifier (QR) that predicts unquantized features from quantized ones using spatial correlations. A soft-to-predictive (STP) training procedure enables seamless integration of QR with existing neural codecs, requiring no major changes to encoders or decoders. Across four baseline architectures, QR yields consistent gains in PSNR and MS-SSIM on Kodak without increasing bitrate, and it can reduce quantization error significantly (up to ~38% in some settings). The method is lightweight, incurring only modest runtime overhead, and it broadens the practical impact of neural compression by reliably improving image quality with minimal computational cost.

Abstract

Neural image compression has been shown to outperform traditional image codecs in terms of rate-distortion performance. However, quantization introduces errors in the compression process, which can degrade the quality of the compressed image. Existing approaches address the train-test mismatch problem incurred during quantization, the random impact of quantization on the expressiveness of image features is still unsolved. This paper presents a novel quantization rectifier (QR) method for image compression that leverages image feature correlation to mitigate the impact of quantization. Our method designs a neural network architecture that predicts unquantized features from the quantized ones, preserving feature expressiveness for better image reconstruction quality. We develop a soft-to-predictive training technique to integrate QR into existing neural image codecs. In evaluation, we integrate QR into state-of-the-art neural image codecs and compare enhanced models and baselines on the widely-used Kodak benchmark. The results show consistent coding efficiency improvement by QR with a negligible increase in the running time.

Neural Image Compression with Quantization Rectifier

TL;DR

This work tackles the quantization-induced loss of feature expressiveness in neural image compression by introducing a Quantization Rectifier (QR) that predicts unquantized features from quantized ones using spatial correlations. A soft-to-predictive (STP) training procedure enables seamless integration of QR with existing neural codecs, requiring no major changes to encoders or decoders. Across four baseline architectures, QR yields consistent gains in PSNR and MS-SSIM on Kodak without increasing bitrate, and it can reduce quantization error significantly (up to ~38% in some settings). The method is lightweight, incurring only modest runtime overhead, and it broadens the practical impact of neural compression by reliably improving image quality with minimal computational cost.

Abstract

Neural image compression has been shown to outperform traditional image codecs in terms of rate-distortion performance. However, quantization introduces errors in the compression process, which can degrade the quality of the compressed image. Existing approaches address the train-test mismatch problem incurred during quantization, the random impact of quantization on the expressiveness of image features is still unsolved. This paper presents a novel quantization rectifier (QR) method for image compression that leverages image feature correlation to mitigate the impact of quantization. Our method designs a neural network architecture that predicts unquantized features from the quantized ones, preserving feature expressiveness for better image reconstruction quality. We develop a soft-to-predictive training technique to integrate QR into existing neural image codecs. In evaluation, we integrate QR into state-of-the-art neural image codecs and compare enhanced models and baselines on the widely-used Kodak benchmark. The results show consistent coding efficiency improvement by QR with a negligible increase in the running time.
Paper Structure (16 sections, 8 equations, 7 figures, 3 tables)

This paper contains 16 sections, 8 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Proposed Method
  4. Formulation of Learned Compression Models
  5. Quantization Rectifier
  6. Soft-to-predictive Training
  7. Experiments
  8. Experimental Setup
  9. Coding Efficiency Improvement
  10. Processing Speed
  11. Conclusion
  12. Detailed Quantization Rectifier Architecture
  13. Learning Coeeficient Exploration in Training
  14. Training Configuration and Detailed Experimental Metrics
  15. Numerical Results of Coding Efficiency Improvement
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Quantization Rectifier Architecture.
  • Figure 2: Coding efficiency of baseline models and their enhanced versions by QR in terms of PSNR and MS-SSIM.
  • Figure 3: Detailed Architecture of the Quantization Rectifier.
  • Figure 4: The impact of rectifier learning coefficient on image quality.
  • Figure 5: Quantization error reduction.
  • ...and 2 more figures