Abstract
<title>Abstract</title> <p>Multi-task image coding for human perception and machine vision has recently attracted increasing attention within unified compression frameworks. However, existing unified coding methods based on shared latent representations still struggle to jointly preserve perceptual texture fidelity and classification-relevant semantic discriminability, especially at low bitrates where generic decoding paths tend to produce over-smoothed reconstructions and attenuated task-relevant semantic cues. To address this issue, we propose a task-adaptive decoder specialization framework for unified image coding toward human reconstruction and machine classification. Built within an all-in-one coding architecture, our framework preserves a shared encoder, entropy model, latent bitstream, and principal decoder backbone, while constructing task-adaptive dynamic decoding paths on the decoder side. This enables reconfiguration of decoding pathways under a unified representation to accommodate diverse task requirements. Specifically, we design two conditionally activated decoder-side adapters. First, the Latent Space Semantic Preservation (LSSP) mechanism dynamically reshapes quantized latents at the decoder input and introduces direct semantic supervision to enhance classification-relevant latent discriminability under low-bitrate compression. Second, the Hierarchical High-Frequency Restoration (HFR) network leverages non-local contextual modeling to compensate for high-frequency residuals at the decoder output without introducing additional transmitted bits, thereby alleviating texture smoothing in generic reconstruction paths. We further employ a two-stage optimization strategy, where the shared codec backbone is first trained and then frozen, while only decoder-side paths and adapter parameters are fine-tuned. Experiments demonstrate that the proposed method outperforms representative state-of-the-art (SOTA) unified coding methods in both low-bitrate machine classification and human-oriented reconstruction quality.</p>