Semantic and Saliency-Aware Scalable Image Coding Toward Human–Machine Collaboration

With the widespread deployment of intelligent vision applications, a substantial amount of visual data is being transmitted to machines for automated analysis to alleviate the human burden. However, research on the interaction between human and machine vision remains limited, significantly constraining the collaborative potential of both systems. In this paper, we investigate task-oriented high- and low-level representations, and how they can be used to construct a scalable coding model for human-machine collaboration. First, we propose a semantic-aware base layer combined with an implicit semantic module, designed to encourage the network to learn compact representations for machine tasks under semantic consistency constraints. Second, we propose a saliency-aware enhancement layer, which navigates the compression of visual signals with a saliency prior derived from the base layer so as to construct high-quality visual perceptions that fit the collaborative scene. By interacting and recombining the decoupled features, our model further bridges the gap between high- and low-level representations, so that the learned representations enjoy both machine analysis and visual reconstruction. Experimental results demonstrate that the proposed method outperforms state-of-the-art machine vision codecs on several machine vision tasks, and can also achieve comparable or even better reconstruction quality, while maintaining a modest bit rate cost.