MoECodec: Image Compression for joint human and machine perception via Mixture-of-Experts

MoECodec is a novel image compression framework that integrates a token-aware Mixture-of-Experts (MoE) approach to optimize performance across multiple vision tasks. By replacing traditional feed-forward network layers in transformer-based models with token-wise MoE, it allows for dynamic computation based on the semantic importance of different image regions. The framework employs a stable routing strategy and a lightweight expert architecture, Group Shuffle MLP (GShMLP), resulting in improved efficiency and effectiveness in both image reconstruction and machine perception tasks, which is crucial for practitioners aiming to enhance model adaptability and performance in diverse applications.