A spatially adaptive compressive sensing framework via joint perception–innovation modeling

Compressive sensing (CS) enables efficient image acquisition with reduced measurements; however, its performance critically depends on how measurements are spatially allocated. Existing methods typically rely on fixed or heuristic adaptive sampling strategies that are not explicitly aligned with the spatial distribution of reconstruction error, resulting in suboptimal measurement utilisation. In this paper, we present a spatially adaptive CS framework driven by a joint perception–innovation metric (JPIM), which acts as a learnable spatial allocation policy that integrates perceptual saliency and reconstruction instability to explicitly guide measurement distribution. Specifically, we observe that effective sampling should prioritise regions that are both structurally informative and difficult to reconstruct, motivating a unified formulation of JPIM that integrates perceptual saliency and reconstruction instability to guide adaptive measurement allocation. Building upon this metric, we develop a feedback-driven adaptive sampling mechanism coupled with a reconstruction subnet that iteratively refines structures and reduces instability. By explicitly modeling the interplay between structural importance and reconstruction instability, the proposed method enables more efficient utilisation of sampling resources compared with conventional heuristic strategies. Extensive experiments on benchmark datasets demonstrate that the proposed approach consistently outperforms both non-adaptive and adaptive CS methods. At a sampling rate of 0.04, the proposed method improves the PSNR/SSIM by 0.14 dB/0.0123 over Uformer-ICS on CBSD68, and by 1.30 dB/0.0215 over MB-RACS on Urban100. These results demonstrate the effectiveness of the proposed JPIM-guided adaptive sampling framework in improving reconstruction fidelity and recovering fine structural details under limited measurement budgets.