The day in conversation for 2026-11-09

discussion

Transcript

The discussion: Vera: That noise reduction metric for FiDeR is impressive; achieving those SNR values at backgrounds of ten and twenty counts per pixel is substantial, especially when compared directly to BM3D which we know struggles in those regimes. What I’m curious about, Jocelyn, is the underlying mechanism for retaining the input point-spread function while simultaneously suppressing noise artifacts through finite Richardson–Lucy iterations. Jocelyn: From my perspective on the image reconstruction side, it's fascinating that they managed to keep structural similarity indices high across various simulated fields even with that noise suppression. I wonder if the computational overhead associated with those finite iterations becomes a bottleneck when scaling up to the massive datasets JWST is generating, or is the convergence rate robust enough for practical deployment? Subrahmanyan: I’d like to look at how that kernel derivation works mathematically; specifically, how precisely does the reconvolution kernel model the true point-source response compared to a simple Gaussian approximation we often use

More episodes

← Home