XPAR-Net: cross-paradigm alignment and reconstruction network for high-fidelity lens-free on-chip microscopy

Lens-free on-chip microscopy (LFOM) enables compact, wide-field imaging, yet reconstructing a complex field from a single in-line hologram involves two coupled challenges: phase must be inferred from an intensity-only measurement, while fringe-scale information must be recovered within a finite numerical-aperture bandwidth. We introduce XPAR-Net, which addresses these challenges through coordinated learning across complementary representation paradigms. Field-wide channel-relation modeling captures selective dependencies among holographic responses, whereas localized spatial refinement preserves fine structural variations. Cross-Domain Alignment and Fusion reconciles these representations and confines learnable spectral residual correction to the modeled acquisition support. Training further couples reference-domain complex-field fidelity with in-support spectral-profile, local-structural, and input-relative refinement constraints. On the complete 11-category in-distribution test set, XPAR-Net improved average PSNR by 1.28 dB and SSIM by $$2.9%$$ over HoloFormer, with only $$8.8%$$ of HoloFormer’s trainable parameters and $$19.0%$$ of its MACs for a 256× 256 input patch. These results highlight coordinated cross-paradigm modeling as an effective route to accurate and computationally compact single-frame LFOM reconstruction under the evaluated conditions.