Integrated Forward–Inverse Network for Physics-Guided Image Reconstruction

ICLR 2026 Conference Submission16203 Authors

Keywords: Computational Imaging, Lensless Imaging, Physics-guided Learning, Inverse Problem

TL;DR: We propose IFIN, the network couples forward physics and learned inverse at every layer with learnable calibration-free PSF, and shows state-of-the-art lensless imaging results under spatially varying blur and noise.

Abstract:

Inverse modeling plays a central role across computational optical imaging problems, including microscopy, imaging through scattering media, and lensless cameras, where the forward model often manifests as a severe blur. Discrepancies between the model and the actual imaging process further aggravate the ill-posed nature of the inverse problem. Physics-enabled methods that integrate analytical forward models with data-driven networks have been explored, but most incorporate physics only in a one-sided manner—either operating purely in the measurement space or only after inversion—thereby discarding complementary cues and reducing robustness to calibration errors.
Here, we propose the Integrated Forward–Inverse Network (IFIN), a physics-guided deep neural network that interleaves differentiable forward operators with learnable inverse modules at every stage of the hierarchy. This design preserves physical consistency while shaping richer feature representations by jointly leveraging information from both measurement and image domains. A physics-guided kernel adaptation further compensates for inaccurate or unavailable PSF calibration, dynamically refining the kernel for blind deconvolution under system constraints.
IFIN is especially effective when measurements are severely blurred by large point-spread functions, where conventional CNN-based inversion is limited by local receptive fields and underutilizes the measurement signal. On challenging lensless imaging benchmarks—including our newly introduced dataset, IFIN achieves state-of-the-art reconstruction quality and improved robustness under noise and model mismatch.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 16203