Two-Stage Octave Residual Network for End-to-End Image Compression

Fangdong Chen, Yumeng Xu, Li WangAAAI 2022 · DOI

Abstract

Recent advancements in learned image compression have leveraged Octave Convolutions (OctConv) to process high- and low-frequency components separately. However, conventional OctConv often suffers from information loss due to the simultaneous execution of inter-frequency communication and spatial down-sampling. This paper introduces the Two-stage Octave Residual (ToRes) network, which decouples these operations to preserve spatial details. Furthermore, a Context Transfer Module is proposed to exploit cross-frequency redundancies during entropy coding. Experimental results demonstrate that the ToRes network outperforms the VVC (H.266) standard by approximately 0.5 dB in PSNR on the Kodak dataset while maintaining lower computational complexity than previous state-of-the-art neural models.

Introduction

Image compression requires balancing the preservation of high-frequency details (textures, edges) with the efficient representation of low-frequency gradients. While Octave Convolutions aim to reduce spatial redundancy by processing these components at different resolutions, they face two primary challenges:

  1. Information Loss: Standard OctConv mixes frequency communication with resolution reduction, leading to data loss critical for high-quality reconstruction.
  2. Underutilized Redundancy: Existing models often fail to exploit the strong correlation between high-frequency structures and their corresponding low-frequency backgrounds.

Proposed Method: The ToRes Architecture

Two-Stage Octave Residual (ToRes) Block

The core innovation is the ToRes block, which splits the processing into two distinct stages to avoid the pitfalls of standard OctConv.

The proposed ToRes block architecture, decoupling inter-frequency communication from spatial sampling.
  • Stage 1 (Communication): High ($X^H$) and low ($X^L$) frequency features exchange information at their original resolutions. This ensures that spatial details are preserved during the interaction.
  • Stage 2 (Sampling): Following communication, the network performs spatial sampling (down-sampling in the encoder, up-sampling in the decoder) using strided vanilla convolutions. This separation allows kernels to learn more specialized features for each task.

Context Transfer Module

To further optimize the bit-rate, the authors introduce a Context Transfer Module within a Variational Autoencoder (VAE) framework. By encoding high-frequency latents first, the model can use them as context to predict the distribution of low-frequency data. This approach effectively reduces cross-frequency redundancy, providing a bit-rate reduction of approximately 10%.

Experimental Results

The model was trained on ImageNet and evaluated on the Kodak Lossless Image Suite and CLIC 2018 datasets.

Rate-distortion performance comparison on Kodak (a, b) and CLIC 2018 (c, d) datasets.

As shown in the rate-distortion curves, the ToRes network consistently outperforms traditional codecs like HEVC and the latest VVC (VTM 10.0). Notably, it achieves a gain of ~0.5 dB PSNR over VVC on the Kodak benchmark. Qualitatively, the model excels at preserving fine textures that are often smoothed out by traditional block-based codecs.

Conclusion

The Two-stage Octave Residual network sets a new state-of-the-art for end-to-end image compression by rethinking frequency-based feature processing. By decoupling communication from sampling and implementing cross-frequency context transfer, the model achieves superior rate-distortion performance with competitive computational efficiency. Future work will focus on further refining entropy models and exploring hardware-friendly implementations for real-time deployment.

Summary updated

BibTeX
@inproceedings{chen2022two,
  author = {Fangdong Chen and Yumeng Xu and Li Wang},
  title = {{Two-Stage Octave Residual Network for End-to-End Image Compression}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i4.20308},
}
Figure