LCS: A Learnlet-Based Sparse Framework for Blind Source Separation

Technical Summary: LCS: A Learnlet-Based Sparse Framework for Blind Source Separation

Problem Statement
Blind Source Separation (BSS) is a fundamental challenge in observational cosmology and astrophysics, tasked with extracting scientifically meaningful signals (such as the Cosmic Microwave Background or HI intensity mapping) from multi-frequency observations contaminated by foregrounds and instrumental noise. Traditional methods, including Principal Component Analysis (PCA), Independent Component Analysis (ICA), and sparse techniques like Generalised Morphological Component Analysis (GMCA), rely on fixed transform domains (e.g., wavelets, starlets) to enforce sparsity. While effective, these fixed-basis approaches may fail to optimally capture the statistical structure of complex, non-Gaussian, or non-stationary astrophysical signals. Conversely, purely data-driven deep learning approaches often lack interpretability and theoretical guarantees regarding signal reconstruction. The paper addresses the need for a hybrid framework that combines the interpretability and sparsity-promoting properties of classical signal processing with the adaptability of deep learning.

Methodology
The authors propose the Learnlet Component Separator (LCS), a novel BSS framework that integrates a learned sparse representation, termed the Learnlet transform, into an iterative source separation loop.

• The Learnlet Transform: This is a structured Convolutional Neural Network (CNN) designed to emulate wavelet-like multiscale behavior while learning filters and thresholding parameters directly from data. The architecture begins with a starlet transform to decompose an image into $J$ scales. The first $J-1$ scales pass through an analysis layer (learned filters), a hard-thresholding layer (learned thresholds), and a synthesis layer (learned filters). The final scale is preserved. Crucially, the synthesis layer is constructed to ensure perfect reconstruction in the absence of noise. The Learnlet network was pre-trained on the ImageNet dataset for image denoising, demonstrating performance comparable to UNet architectures but with significantly fewer parameters (~12,800 vs. millions) and exact reconstruction properties.
• The LCS Algorithm: LCS operates as an iterative algorithm inspired by GMCA but with distinct modifications:
  1. Source Estimation: Given a fixed mixing matrix estimate, sources are estimated by minimizing a cost function that includes a data fidelity term and a sparsity term. The sparsity is enforced via the Learnlet transform using hard-thresholding (avoiding the bias of soft-thresholding used in GMCA).
  2. Mixing Matrix Estimation: The mixing matrix is refined by minimizing the error between the observed data (excluding coarse scales to remove low-frequency correlations) and the product of the mixing matrix and the sparse source estimates.
  3. Post-Processing: A final denoising step is applied to the estimated sources using the Learnlet transform with noise levels derived from the estimated source covariance.
  4. Training Strategy: The framework supports two modes: using a shared, general-purpose basis (pre-trained on ImageNet) or individually retraining specific Learnlet bases ($L_i$) for each source component to optimize sparsity for specific morphologies.

Key Contributions

1. Hybrid Framework: The introduction of LCS, which bridges classical sparse BSS methods and deep learning by embedding a learned, wavelet-inspired transform into an iterative separation loop.
2. Learnlet Integration: The adaptation of the Learnlet transform—originally designed for denoising—into a structured component separator that preserves multiscale structure and interpretability while gaining data-driven adaptability.
3. Algorithmic Improvements over GMCA: LCS introduces hard-thresholding (reducing bias), updates sources in real space rather than the wavelet domain, and includes a dedicated post-processing denoising step.
4. Flexible Training: The demonstration that LCS can function effectively with a general pre-trained basis or be fine-tuned with component-specific bases to handle diverse signal morphologies.

Experimental Results
The authors evaluated LCS on four distinct experiments involving both synthetic "toy" models and real astrophysical data:

• Standard Test Images (Common Basis): Using images of "Barbara," an airplane, and a boat, LCS outperformed GMCA (starlet and biorthogonal variants) and FastICA. In low-noise regimes, LCS showed a 3–5 dB gain in Peak Signal-to-Noise Ratio (PSNR); in high-noise regimes, gains reached 10–15 dB. LCS also achieved superior mixing matrix reconstruction (5–15 dB gain in Component Angle metric).
• Astrophysical X-ray Data (Common Basis): Compared against LPALM (a semi-supervised deep learning method) on supernova remnant data, LCS performed comparably at high Signal-to-Noise Ratios (SNR) but significantly outperformed LPALM as noise increased. Notably, LCS remained robust when the mixing matrix was perturbed, whereas LPALM's performance degraded, highlighting LCS's advantage as a fully blind method.
• Texture Data (Individual Bases): Using the Describable Texture Dataset, individually retrained Learnlet bases further enhanced separation performance compared to a common basis, particularly in mixing matrix estimation.
• CMB and SZ Extraction: Applied to WebSky simulations of CMB and Sunyaev-Zel'dovich (SZ) signals, LCS achieved the highest PSNR for both components (34.65 dB for CMB and 42.94 dB for SZ), outperforming ILC and GMCA. The results showed effective removal of SZ contamination from CMB maps and vice versa.

Significance and Claims
The paper claims that LCS represents a significant step forward in handling the complexity of next-generation cosmological experiments, such as the Square Kilometre Array (SKA). The authors emphasize that LCS offers a "flexible, interpretable, and high-performing approach" that overcomes the limitations of fixed-basis methods without sacrificing the theoretical grounding of sparse representation.

The results highlight that hybrid approaches, combining signal processing priors with deep learning, can achieve superior separation performance, particularly in scenarios involving non-stationary foregrounds or non-ideal mixing conditions. The authors note that while LCS provides visually and quantitatively accurate reconstructions, it does not inherently enforce the unbiased residual constraint found in Internal Linear Combination (ILC) methods, suggesting that future work may require explicit debiasing steps for cross-correlation analyses. The study positions LCS as a promising tool for extracting faint cosmological signals (like HI intensity mapping) from bright foregrounds, leveraging the robustness of learned sparse representations.