Interference data denoising and baseline correction via a multi-scale deep denoising network and multidimensional gradient-consistent regularization

Abstract

The complex process of spectral inversion often results in the degradation of essential information, with noise and baseline signals significantly impacting the precision and accuracy of material analyses. Recent deep learning-based noise and baseline removal approaches exhibit better performance, however, the high spectral resolution characteristics of spatial heterodyne interference data and the scarcity of labeled data still restrict extensive applications and developments of deep learning in this field. To address these issues, we incorporate the signal generation mechanism (e.g., statistical analyses of interference noise and baseline signals) of spatial heterodyne spectroscopy systems to construct a mathematical signal simulation model and corresponding simulation database, which can contribute to circumventing the problem of data scarcity.In light of this, we propose a multi-level signal enhancement network model (InDNet) based on multi-scale local information extraction and transformer-based global information fusion, which simultaneously achieves denoising and baseline correction. Notably, by virtue of the spatial features of the interference pattern, we designed a Multidimensional Gradient-consistent Regularization to supervise the training process of the network. Extensive experiments demonstrate that our method achieves excellent results on both simulated and real datasets (SSIM values of 0.9757 and 0.9827, respectively), significantly outperforming methods such as PDNet and LRDUNet. Our approach effectively accomplishes both denoising and baseline correction tasks, indirectly confirming the validity of the interference data simulation model. This study provides insights into the intelligent processing of spatial heterodyne interference data and offers new ideas for enhancing the quality of spectral signals.