Multiscale Photon-limited spectral image reconstruction

This paper studies photon-limited spectral intensity estimation and proposes a spatially and spectrally adaptive, nonparametric method for estimating spectral intensities from Poisson observations. Specifically, our method searches through estimates defined over a family of recursive dyadic partitions in both the spatial and spectral domains, and finds the one that maximizes a penalized log likelihood criterion. The key feature of this approach is that the partition cells are anisotropic across the spatial and spectral dimensions, so that the method adapts to varying degrees of spatial and spectral smoothness, even when the respective degrees of smoothness are not known a priori. The proposed approach is based on the key insight that spatial boundaries and singularities exist in the same locations in every spectral band, even though the contrast or perceptibility of these features may be very low in some bands. The incorporation of this model into the reconstruction results in significant performance gains. Furthermore, for spectral intensities that belong to the anisotropic Ḧolder-Besov function class, the proposed approach is shown to be near-minimax optimal. The upper bounds on the risk function, which is the expected squared Hellinger distance between the true intensity and the estimate obtained using the proposed approach, matches the best possible lower bound up to a log factor for certain degrees of spatial and spectral smoothness. Experiments conducted on realistic data sets show that the proposed method can reconstruct the spatial and the spectral inhomogeneities very well even when the observations are extremely photon-limited (i.e., less than 0.1 photon per voxel).