MT-MoA: A Unified Parameter-Efficient Multi-Task Adaptation Framework Toward Image Coding for Machines

With the increasing demand for machine vision applications, Image Coding for Machines (ICM) has garnered increasing attention, focusing on compressing images to support downstream analytics rather than human perceptual fidelity. Among ICM methods, transfer-based methods aim to adapt pre-trained human-oriented codecs to individual tasks with minimal parameter overhead. However, they suffer from linearly increasing computational costs and underutilization of shared information across tasks. To address these issues, we propose MT-MoA, a unified parameter-efficient multi-task adaptation framework for ICM. Inspired by recent advances in Mixture-of-Experts (MoE), we propose Instance-Aware Mixture-of-Adapters (IA-MoA) modules in the encoder to capture task-agnostic features, and Task-Aware Mixture-of-Adapters (TA-MoA) modules in the decoder to perform task-conditioned reconstruction. To achieve high parameter efficiency and promote cross-task information transfer, we propose a shared adapter pool that learns transferable features across instances and tasks, while a dynamic router that allocates adapters adaptively based on input or task requirements. Extensive experiments on PASCAL-Context demonstrate that MT-MOA achieves state-of-the-art rate-accuracy performance across diverse machine tasks. Further evaluation on MS-COCO across additional tasks validates its strong scalability and generalization capability.