Physics-guided multi-exposure fusion pushes hyperspectral imaging into HDR scenes
ssh4net · x · 2026-11-13
A new paper tackles high-dynamic-range (HDR) scenes in computational hyperspectral imaging.
Problem: encoding-mask channels have unequal spectral responsivities, so each channel hits its reliability limit at different signal levels — a single exposure can't preserve spectral info in both highlights and shadows.
Method: a physics-guided multi-exposure fusion framework. Full-field sensor-response characterization yields channel-aware reliability gates; each exposure passes through a shared reliability estimator (analytical physical prior + learned adjustment head) producing spatially and channel-varying weights, fused symmetrically in the encoded measurement domain, then refined by a residual network before reconstructing a 31-band, 400–700nm hyperspectral cube.
Results: on a real system with multi-exposure ColorChecker captures (7.8/15.6/62.5ms), reconstructed spectra closely match a spectrometer reference; the method achieves lower average SAM error (0.0285 rad vs 0.0410 rad single-exposure) than Hat fusion, SCTNet and AHDRNet, with PSNR improving as more exposures are added.
More from Research
- Question's Gambit: first-step reranking lifts agentic search accuracy to 90.5% — CShorten30 · 2026-11-13
- Terence Tao: major Riemann hypothesis progress came from AI-generated ideas nobody yet understands — burny_tech · 2026-11-13
- Deleting under 0.05% of weights matches full fine-tuning, NYU paper shows — burny_tech · 2026-11-13
- Recasting spatial-spectral coupling as attention scales 3D shape matching to high-res meshes — ssh4net · 2026-11-13
- Federal Reserve releases FEDS paper on AI's current impact on the U.S. economy — asusarla · 2026-11-13
- Making LLM inference bit-identical across GPUs and CPUs with integer-only arithmetic — CatAstro_Piyush · 2026-11-13