Learning Binary Sampling Patterns for Single-Pixel Imaging using Bilevel Optimisation

2025-08-26

AI Summary

To address the reconstruction performance bottleneck under highly undersampled conditions in single-pixel imaging, this paper proposes a task-specific binary illumination pattern learning framework. We formulate a bilevel optimization scheme that jointly optimizes binary sampling patterns and a deep reconstruction network: the upper-level objective maximizes reconstruction fidelity, while the lower-level performs end-to-end image recovery. To handle the non-differentiability of binary constraints, we incorporate a straight-through estimator (STE); additionally, total deep variation (TDV) regularization is embedded to enhance reconstruction robustness. Experiments on the CytoImageNet fluorescence microscopy dataset demonstrate that the learned binary patterns significantly outperform conventional random and Hadamard patterns—as well as state-of-the-art learned methods—at extremely low sampling rates (e.g., 5%), achieving an average PSNR gain of 2.1 dB. These results validate the effectiveness and generalizability of co-optimizing illumination patterns and reconstruction networks.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Constraint Learning and Acquisition

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Security and Privacy: Large-scale security measurements

Abstract

Single-Pixel Imaging enables reconstructing objects using a single detector through sequential illuminations with structured light patterns. We propose a bilevel optimisation method for learning task-specific, binary illumination patterns, optimised for applications like single-pixel fluorescence microscopy. We address the non-differentiable nature of binary pattern optimisation using the Straight-Through Estimator and leveraging a Total Deep Variation regulariser in the bilevel formulation. We demonstrate our method on the CytoImageNet microscopy dataset and show that learned patterns achieve superior reconstruction performance compared to baseline methods, especially in highly undersampled regimes.

Problem

• Learning binary illumination patterns for single-pixel imaging
• Optimizing task-specific patterns using bilevel optimization
• Overcoming non-differentiable binary pattern optimization challenges

Innovation

• Bilevel optimisation for binary patterns
• Straight-Through Estimator handles non-differentiability
• Total Deep Variation regulariser in formulation