MoECodec: Image Compression for joint human and machine perception via Mixture-of-Experts
MoECodec introduces a unified image compression framework using Mixture-of-Experts (MoE) routing to dynamically adapt compression based on image content and downstream vision tasks. The approach reduces computational overhead compared to task-specific models while maintaining performance across multiple machine perception applications.
MoECodec addresses a fundamental challenge in machine vision: creating efficient compression systems that serve multiple downstream tasks without requiring separate models for each application. Traditional approaches either demand task-specific architectures with substantial parameter overhead or rely on static compression methods that treat all image regions equally despite their varying semantic importance. This research shifts compression from a one-size-fits-all paradigm to a token-aware system where computational resources are allocated dynamically based on content complexity and task requirements.
The innovation centers on replacing standard feed-forward network layers with token-wise Mixture-of-Experts routing, enabling the model to assign different computational paths to different image regions. To stabilize this approach, the authors combine expert-choice routing with spatial total variation regularization, ensuring that adjacent image regions receive similar treatment. The introduction of Group Shuffle MLP (GShMLP) keeps parameter growth manageable, critical for practical deployment. This represents an evolution in efficient AI architecture design, where routing mechanisms enable more intelligent resource allocation than fixed computation patterns.
For the broader AI infrastructure ecosystem, unified compression codecs reduce deployment complexity and memory footprint—significant considerations for edge devices and large-scale vision systems. Fewer model variants simplify maintenance and reduce inference latency, directly benefiting companies operating vision pipelines. The technique demonstrates how modern architectural innovations like MoE can extend beyond language models into specialized domains like compression. As vision tasks proliferate across industries, efficient multi-task compression becomes increasingly valuable.
- →MoECodec enables dynamic token-level computation that adapts compression based on image content and task requirements rather than applying static transformations
- →The stable routing strategy combining expert-choice routing with spatial regularization prevents fragmented expert assignments and improves performance consistency
- →Single unified model replaces multiple task-specific compression architectures, reducing deployment overhead and parameter count
- →The framework shows consistent improvements across both image reconstruction and multiple downstream machine vision tasks in experiments
- →Group Shuffle MLP architecture controls parameter growth while maintaining the benefits of dynamic expert routing