GLIC: General Format Learned Image Compression

Mingsheng Zhou, Mingming KongAAAI 2025 · DOI

In the landscape of deep learning-based image compression, most state-of-the-art models are specialized for three-channel RGB formats. This specialization creates significant hurdles when processing diverse data types, such as grayscale medical X-rays or multi-channel satellite imagery. Researchers from XiHua University have proposed GLIC: General Format Learned Image Compression, a framework that unifies any image format into a single-channel representation, achieving superior rate-distortion performance and enabling progressive decoding.

The Problem: Format Specialization

Current Learned Image Compression (LIC) models are often "hardcoded" for $C=3$ channels. This leads to several issues:

  1. Format Diversity: Industrial and medical sensors produce data with varying channel counts (e.g., CT/MRI scans or spectral data).
  2. Coding Efficiency: Standard multi-channel convolutions can interfere with "homomorphism"—the local consistency of pixels—making data harder for entropy models to compress.
  3. Progressive Decoding: Most existing architectures do not natively support high-quality "preview" decoding during bitstream transmission.

Methodology

GLIC introduces three core innovations to create a universal compressor:

1. One-Channel Unification

Instead of processing images as $C \times H \times W$, GLIC utilizes a Reconstruction Split and Concat (SC/RSC) module. This module tiles multiple channels into a single, larger two-dimensional matrix (e.g., an RGB image becomes a $1 \times 3H \times W$ matrix). This ensures the neural network treats all data types as a unified 2D format.

2. Adaptive Attention Residual Block (AARB)

To generalize across different visual data types, the AARB employs a dual-pathway architecture. An Attention Branch extracts high-level semantic information, while a Non-Attention Branch retains low-level details. A dynamic module adaptively weights these paths based on the input features.

3. Evenly Grouped Cross-Channel Context (CCCT)

For entropy coding, GLIC splits the latent representation into eight even groups. This design choice facilitates progressive decoding, allowing the decoder to generate a low-quality preview from the first few groups and refine it as more bits arrive.

Experimental Results

GLIC was evaluated on a diverse set of benchmarks, including the Kodak dataset (natural images), LUNA16 (medical lung nodules), and Bluetooth RF spectral data.

Rate-distortion performance on the Kodak dataset. The inset highlights the performance gains attributed to the Adaptive Attention Residual Block (AARB) and the One-Channel unification strategy.

As shown in the figure above, GLIC outperforms existing state-of-the-art models like ELIC and Cheng2020 on the Kodak dataset. Notably, the model trained on natural images effectively compressed Radio Frequency (RF) data, demonstrating the robustness of the "one-channel" philosophy.

Analysis: The Homogeneity Advantage

A key insight of the GLIC framework is that tiling channels into a single plane preserves local textures better than traditional multi-channel processing. The researchers found that while information entropy remains similar across models, the Homogeneity Index is significantly higher in the one-channel approach.

Comparison of Entropy and Homogeneity across different LIC models. GLIC ("Ours") demonstrates a significantly higher Homogeneity Index, which facilitates more efficient entropy coding.

This increased homogeneity makes the entropy coder's task more efficient, directly contributing to the model's superior rate-distortion performance. Furthermore, the use of even grouping in the CCCT module proved superior to the uneven grouping used in models like ELIC for maintaining preview quality during progressive decoding.

Conclusion

GLIC represents a shift toward more flexible and efficient image compression. By treating all images as a general 2D format and optimizing for local homogeneity, the framework achieves state-of-the-art results across natural, medical, and industrial data types. Future work will explore more complex flattening operations to further optimize the mapping of multi-dimensional data to the 2D plane.

Summary updated

BibTeX
@inproceedings{zhou2025glic,
  author = {Mingsheng Zhou and Mingming Kong},
  title = {{GLIC: General Format Learned Image Compression}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i10.33175},
}
Figure