.png){kind=icon}
.png){kind=icon}
Return
Physics-Consistent Imaging Inverse Problem Solving Framework for Heterogeneous Systems via Latent Space Optimization
.png){kind=icon}
PRE
AI
DOI：10.1109/TCI.2026.3675456.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Modern imaging systems face diverse degradation artifacts arising from complex sensing mechanisms and noise environments. Existing methods often require extensive parameter tuning or retraining across different systems and degradation levels. Although recent generative-prior based approaches partially alleviate this issue, they typically assume that target scenes lie within the pretrained model’s manifold; otherwise, optimization may yield degenerate solutions lacking target features and structural integrity. We propose a physics-guided framework for solving imaging inverse problems in heterogeneous systems. The framework reconstructs scenes progressively, from coarse structures to fine details, through a three-phase procedure involving latent-vector optimization and latent-space fine-tuning, thereby relaxing the manifold constraint on target scenes. Furthermore, we design a prior module based on a local linearity assumption in the latent space, which improves reconstruction speed and quality for image pairs with structural similarity through warm start initialization. Comprehensive experiments on CMOS sensors, single-photon avalanche diode (SPAD) arrays, and single-pixel imaging systems (SPIS) demonstrate that the proposed framework consistently delivers competitive performance under extreme environmental conditions, ultra-low-bandwidth constraints, and out-of-distribution scenarios. Code is available at Inverse.
Keywords:
Heterogeneous systems
inverse problem
domain generalization
latent space
Journal
I
IF:
4.8
Papers:
127
Citations:
0