💻 computer science

Adaptive Learned Image Compression with Graph Neural Networks

This paper proposes GLIC, a graph neural network-based learned image compression framework that utilizes dual-scale graphs and adaptive connectivity to overcome the rigid receptive fields of CNNs and Transformers, achieving state-of-the-art performance with significant BD-rate reductions on standard benchmarks.

Original authors: Yunuo Chen, Bing He, Zezheng Lyu, Hongwei Hu, Qunshan Gu, Yuan Tian, Guo Lu

Published 2026-03-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yunuo Chen, Bing He, Zezheng Lyu, Hongwei Hu, Qunshan Gu, Yuan Tian, Guo Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to pack a massive, messy attic full of items into a small moving truck. The goal is to fit everything in (compression) without losing anything important (quality).

For years, the best way to do this was like using a standard grid of boxes. You take a 3x3 grid, put it over a section of the attic, and pack whatever is inside. If you move the grid, you do the same thing again.

The Problem: This approach is rigid. It treats a smooth, empty wall the same way it treats a chaotic pile of broken toys. It also only looks at what's immediately next to it, missing the fact that a toy on the far left might be identical to a toy on the far right. It's like trying to describe a whole city by only looking at the block you are standing on, ignoring the fact that the whole city follows a pattern.

The paper you shared introduces a new, smarter way to pack: GLIC (Graph Neural Network-based Image Compression).

Here is how it works, using simple analogies:

1. The "Smart Map" Instead of a Grid

Instead of using a rigid grid, GLIC draws a dynamic map (a graph) over the image.

Old Way: "I will look at the 8 pixels touching this one."
GLIC Way: "I will look at the 8 pixels touching this one, PLUS I will look at a few pixels far away that look exactly like this one."

Think of it like a social network. In a rigid grid, you only talk to your immediate neighbors. In GLIC, if you are a "red feather" in a picture of a parrot, the system instantly connects you to every other red feather in the image, even if they are on the other side of the photo. This allows the system to say, "Oh, I already know what a red feather looks like; I don't need to write down the details for every single one."

2. The "Two-Team" Strategy (Dual-Scale Graphs)

GLIC uses two types of connections simultaneously, like a team with two different roles:

The Local Team (Dense): They look closely at the immediate neighborhood to catch fine details, like the texture of grass or the edge of a leaf. This is like a detective looking at footprints right next to you.
The Global Team (Sparse): They look at the whole picture from a distance to find patterns. This is like a drone flying overhead to see that the whole field is green, so you don't need to describe every single blade of grass.

By combining these, the system gets the best of both worlds: it keeps the sharp details while realizing that big parts of the image are just repeats of each other.

3. The "Busy vs. Chill" Neighborhoods (Adaptive Connectivity)

This is the paper's biggest innovation. Not all parts of an image are equally hard to compress.

Smooth areas (The "Chill" Neighborhood): A blue sky or a white wall has very little detail. GLIC says, "This is easy. Let's just connect a few neighbors and move on." It saves energy here.
Complex areas (The "Busy" Neighborhood): A parrot's feathers or a busy city street are full of detail. GLIC says, "This is hard! Let's connect many neighbors to figure out exactly how to describe this."

The system uses a complexity score (like a "stress meter") to decide how many connections each pixel needs. It's like a traffic controller: it directs more resources to the busy intersections and fewer to the empty country roads.

4. The Result: Smaller Files, Better Pictures

Because GLIC is so smart about where to look and how much effort to spend:

It finds more redundancy: It realizes that the "red feather" on the left is the same as the "red feather" on the right, so it doesn't waste space describing it twice.
It keeps the details: Because it spends extra effort on complex areas, the picture doesn't get blurry or "blocky."

The Bottom Line:
Traditional methods are like a robot that follows a strict checklist, looking at the same small square over and over. GLIC is like a human expert who looks at the whole picture, spots patterns instantly, ignores the boring parts, and focuses all their energy on the interesting parts.

The authors tested this against the current "gold standard" of compression (VTM-9.1) and found that GLIC could shrink files by 19% to 21% more than the competition without losing any visible quality. It's a massive win for saving storage space and speeding up internet transmission.

1. Problem Statement

Current State-of-the-Art (SOTA) Learned Image Compression (LIC) methods primarily rely on Convolutional Neural Networks (CNNs) or Transformers. While effective, these architectures suffer from inherent rigidity:

Fixed Receptive Fields: Standard CNN kernels and window-based attention mechanisms connect pixels based on fixed Euclidean distances (e.g., a $k \times k$ window).
Static Connectivity: They couple pixels simply because they are spatially close, regardless of whether they share semantic or structural redundancy.
Inefficiency in Global Modeling: Natural images exhibit non-uniform redundancy (smooth regions vs. textured edges). Rigid models struggle to adaptively capture spatially varying redundancy, particularly long-range correlations, without incurring the quadratic computational complexity of global attention mechanisms.

The core problem is the inability of existing models to dynamically adjust connectivity based on local content complexity, leading to suboptimal rate-distortion performance.

2. Methodology: GLIC Framework

The authors propose GLIC (Graph-based Learned Image Compression), a content-adaptive framework that replaces rigid convolution/attention blocks with Graph Neural Networks (GNNs). The core innovation is the Graph-based Feature Aggregation (GFA) block, which operates in two main stages of the compression pipeline.

Key Components of GFA:

Dual-Scale Graph Construction (Flexible Receptive Fields):
Instead of a single fixed window, each pixel (node) constructs a candidate set from two sources to balance local detail and global context:
- Dense Local Sampling: A fixed-size neighborhood (similar to CNNs) to capture fine-grained structures.
- Sparse Global Sampling: A mesh-grid sampling scheme (inspired by dilated convolutions) that selects pixels across the entire image at specific strides.
- Result: This creates a "near-full" receptive field with near-linear computational complexity, avoiding the $O(N^2)$ cost of global attention.
Complexity-Aware Degree Budgeting (Adaptive Connectivity):
The number of neighbors (degree) for each node is not fixed but dynamically determined by the local content complexity:
- Scoring: A RMS-Gradient (RMS-G) score is calculated using Sobel operators to measure local spatial variation (edges, textures).
- Allocation: Pixels with high complexity (hard to compress) are assigned a higher "neighbor quota" (more connections), while smooth regions receive fewer connections.
- Graph Construction: For each node, neighbors are selected from the dual-scale candidate set based on cosine similarity thresholds, determined via a fast bisection search to meet the target degree quota.
Edge-Conditioned Aggregation:
Features are updated by aggregating information from selected neighbors, weighted by a learnable function of the edge similarity. This allows the model to selectively fuse relevant features while ignoring irrelevant ones.

3. Key Contributions

Dual-Scale Graph Architecture: A novel mechanism enabling flexible receptive fields where pixels can dynamically select relevant nodes from both local neighborhoods and distant global regions.
Content-Adaptive Connectivity: A complexity-aware scoring mechanism (RMS-Gradient) that dynamically assigns connection degrees per pixel, allocating more computational capacity to structurally complex regions.
Efficient GNN-Based LIC: The GLIC model achieves near-global connectivity with linear complexity ($O(Nn) + O(NK)$), overcoming the quadratic bottleneck of global Transformers.
State-of-the-Art Performance: The model significantly outperforms existing codecs in rate-distortion efficiency.

4. Experimental Results

The authors evaluated GLIC on Kodak, Tecnick, and CLIC datasets, comparing against traditional codecs (VTM-9.1) and SOTA LIC models (FTIC, MambaIC, TCM-L, etc.).

Rate-Distortion Performance:
- GLIC achieved BD-rate reductions of -19.29% (Kodak), -21.69% (Tecnick), and -18.71% (CLIC) relative to VTM-9.1.
- It consistently outperformed other learned methods, showing gains of 0.26–0.56 dB in BD-PSNR over competitors like FTIC and TCM-L.
Computational Efficiency:
- Despite the added complexity of graph construction, GLIC maintains high efficiency.
- Compared to MambaIC (CVPR'25), GLIC reduced parameters by 57.2%, FLOPs by 55.4%, and decoding latency by 41.0%.
- The graph construction overhead (bisection search) adds less than 5% to encoding/decoding time.
Effective Receptive Fields (ERF):
- Visualizations show GLIC has a significantly larger and more adaptive ERF than CNN/Transformer baselines.
- Crucially, the ERF is content-aware: it focuses on semantically similar regions (e.g., parrot feathers) even if they are spatially distant, while ignoring irrelevant nearby pixels (e.g., background grass).

5. Significance

This paper represents a paradigm shift in learned image compression by moving away from rigid, Euclidean-space-based connectivity toward dynamic, data-driven graph structures.

Theoretical Impact: It demonstrates that GNNs can effectively model non-Euclidean relationships in images, capturing long-range dependencies without the prohibitive cost of global attention.
Practical Impact: GLIC offers a highly efficient solution that bridges the gap between the flexibility of global modeling and the efficiency of local processing. It proves that adaptive connectivity based on content complexity is a superior strategy for removing spatial redundancy, leading to superior compression ratios and visual quality.
Future Direction: The work opens new avenues for applying graph-based adaptive mechanisms in other vision tasks where spatial redundancy is non-uniform.