Abstract
One-bit compressive sensing is concerned with the accurate recovery of an underlying sparse signal of interest from its one-bit noisy measurements. The conventional signal recovery approaches for this problem are mainly developed based on the assumption that an exact knowledge of the sensing matrix is available. In this work, however, we present a novel data-driven and model-based methodology that achieves blind recovery; i.e., signal recovery without requiring the knowledge of the sensing matrix. To this end, we make use of the deep unfolding technique and develop a model-driven deep neural architecture which is designed for this specific task. The proposed deep architecture is able to learn an alternative sensing matrix by taking advantage of the underlying unfolded algorithm such that the resulting learned recovery algorithm can accurately and quickly (in terms of the number of iterations) recover the underlying compressed signal of interest from its one-bit noisy measurements. In addition, due to the incorporation of the domain knowledge and the mathematical model of the system into the proposed deep architecture, the resulting network benefits from enhanced interpretability, has a very small number of trainable parameters, and requires very small number of training samples, as compared to the commonly used black-box deep neural network alternatives for the problem at hand.
Introduction
Compressive Sensing (CS) is a powerful framework for capturing sparse signals using far fewer measurements than the Nyquist rate requires. In practical hardware implementations, these measurements are often quantized. "One-bit CS" represents the extreme version of this quantization, where only the sign of the measurement (positive or negative) is recorded. While this significantly simplifies hardware design, it introduces massive non-linear distortion and information loss.
The primary challenge in most one-bit CS algorithms is the requirement for exact knowledge of the sensing matrix ($\Phi$)—the mathematical description of the sampling process. In real-world scenarios, this matrix is often unknown, poorly calibrated, or subject to environmental fluctuations. "Blind" recovery—reconstructing the signal without knowing $\Phi$—is notoriously difficult given the high information loss of one-bit quantization.
Methodology: Deep Unfolding for Blind Recovery
The authors employ Deep Unfolding (or Deep Unrolling) to bridge the gap between iterative algorithms and neural networks. They map the iterations of the classical Binary Iterative Hard Thresholding (BIHT) algorithm onto the layers of a deep neural network.
Model-Driven Architecture
Unlike "black-box" neural networks, every layer in this architecture corresponds to a specific step in the BIHT algorithm: a gradient update followed by a hard thresholding operation ($H_k$) to enforce sparsity. This ensures the network remains interpretable and physically grounded.
Training Strategy
The model utilizes a two-stage training process:
- System Identification: The network learns a "surrogate" sensing matrix ($\Phi^*$) that minimizes reconstruction error across a dataset of input-output pairs.
- Learning to Optimize: With the matrix fixed, the network learns optimal, per-layer step-sizes ($\alpha_i$) to accelerate convergence.
To overcome the non-differentiability of the sign function used in one-bit quantization, the authors utilize a Straight-Through Estimator (STE), allowing gradients to flow through the network during backpropagation.
Experimental Results
The framework was validated using synthetically generated $K$-sparse signals ($n=128$) and a $512 \times 128$ sensing matrix with additive Gaussian noise.
Convergence and Efficiency
One of the most significant advantages of the unfolded model is its convergence speed. The proposed architecture achieves a lower Normalized Mean Square Error (NMSE) in just 10 layers than the traditional BIHT algorithm achieves over many more iterations.
Robustness to Sparsity and Noise
Surprisingly, the proposed blind method often outperforms the standard BIHT algorithm even when the latter has full access to the true sensing matrix. By learning a surrogate matrix optimized for the specific recovery task and noise profile, the network achieves superior accuracy. As the sparsity level ($K$) increases—making the signal harder to recover—the performance of the proposed method degrades much more slowly than the state-of-the-art BIHT baseline.
Conclusion and Future Work
This work demonstrates that deep unfolding can effectively solve the blind recovery problem in one-bit compressive sensing. By incorporating domain knowledge, the resulting network is more efficient, requires less data, and is more interpretable than generic deep learning models.
Future research directions include addressing "doubly blind" scenarios where quantization thresholds are also unknown, and developing adaptive thresholding operators that do not require prior knowledge of the signal's sparsity level $K$. Applying this framework to real-world domains such as dynamic MRI or radar imaging remains a key objective for proving its practical utility.
Summary updated