Introducing a multiscale feature integration network for inpainting with applications to enhanced CMB map reconstruction

TL;DR

SkyReconNet is a novel neural network that combines dilated and standard convolutions to effectively inpaint missing regions in full-sky CMB maps, preserving structural features and achieving near-optimal accuracy limited only by cosmic variance.

Contribution

The paper introduces SkyReconNet, a hybrid convolutional neural network designed specifically for inpainting large, irregular masked regions in CMB maps, with broad applicability to other physics-based data reconstruction tasks.

Findings

Achieves high-fidelity CMB map inpainting close to cosmic variance limits.

Effectively reconstructs large, irregular masked regions in full-sky maps.

Demonstrates versatility for other data with missing or defective pixels.

Abstract

We introduce a novel neural network, SkyReconNet, which combines the expanded receptive fields of dilated convolutional layers along with standard convolutions, to capture both the global and local features for reconstructing the missing information in an image. We implement our network to inpaint the masked regions in a full-sky Cosmic Microwave Background (CMB) map. Inpainting CMB maps is a particularly formidable challenge when dealing with extensive and irregular masks, such as galactic masks which can obscure substantial fractions of the sky. The hybrid design of SkyReconNet leverages the strengths of standard and dilated convolutions to accurately predict CMB fluctuations in the masked regions, by effectively utilizing the information from surrounding unmasked areas. During training, the network optimizes its weights by minimizing a composite loss function that combines the Structural Similarity Index Measure (SSIM) and mean squared error (MSE). SSIM preserves the essential structural features of the CMB, ensuring an accurate and coherent reconstruction of the missing CMB fluctuations, while MSE minimizes the pixel-wise deviations, enhancing the overall accuracy of the predictions. The predicted CMB maps and their corresponding angular power spectra align closely with the targets, achieving the performance limited only by the fundamental uncertainty of cosmic variance. The network's generic architecture enables application to other physics-based challenges involving data with missing or defective pixels, systematic artefacts etc. Our results demonstrate its effectiveness in addressing the challenges posed by large irregular masks, offering a significant inpainting tool not only for CMB analyses but also for image-based experiments across disciplines where such data imperfections are prevalent.