TL;DR
This paper presents a novel compressive sensing (CS) approach for passive millimeter-wave (MMW) imaging using Hadamard transforms. By employing a single-pixel imager and a specially designed Hadamard mask, they demonstrate high-fidelity image reconstruction with significantly fewer measurements than traditional methods, achieving a 9x reduction in acquisition time.
What Problem Does This Paper Address?
Traditional imaging systems, especially at millimeter-wave frequencies, often require a large number of measurements to achieve good spatial resolution, leading to long acquisition times. This is particularly challenging for passive MMW imaging, where the signal-to-noise ratio (SNR) can be low. While compressive sensing (CS) offers a promising solution by acquiring fewer samples than dictated by the Nyquist limit, its implementation in MMW imaging is hindered by the lack of suitable spatial modulators like digital micromirror devices (DMDs) used in optical CS. This paper addresses this gap by proposing and demonstrating a Hadamard transform-based CS technique for passive MMW imaging that overcomes these implementation challenges.
Methods
The core of the proposed method is the use of a Hadamard transform. Instead of random masks used in typical CS, they employ a structured Hadamard mask. The imaging setup utilizes a single-pixel passive MMW imager. An extended 2D Hadamard mask, created by periodically extending a cyclic matrix, is raster-scanned across the scene. For each position, a submask of size $p \times q$ (where $p=41$ and $q=43$) exposes a portion of the scene. The transmitted radiation is then collected by a horn antenna and processed by a radiometer.
To achieve compressive sensing, they don't acquire all possible mask positions ($N=pq$). Instead, they use a subset of these measurements. To reconstruct the full image from this sub-sampled data, they developed a "relaxation technique." This iterative method estimates the missing Hadamard measurements by calculating a weighted average of surrounding known measurements, effectively filling in the gaps in the Hadamard measurement space before performing the final image reconstruction using a matrix operation.
Datasets
Not discussed in the paper. The authors used a controlled laboratory setup with an incandescent lamp as an illumination source and a target object (a circular hole with a metal strip) for imaging.
Key Insights and Ideas
- Hadamard Masks for MMW CS: The paper demonstrates that structured Hadamard masks, rather than random ones, can be effectively used for compressive sensing in MMW imaging, circumventing the need for complex spatial light modulators.
- Relaxation Technique for Sub-sampled Hadamard Data: The developed iterative relaxation method is crucial for recovering the full Hadamard measurement space from sparse acquisitions, enabling high-fidelity reconstruction.
- Trade-off in SNR for Resolution: While using subwavelength pixels (1.24 mm for 2 mm wavelength) with the Hadamard mask leads to a reduction in SNR (from 13.1 to 7.9), this is deemed an acceptable trade-off for achieving higher image resolution.
- Real-time Reconstruction Potential: The inherent structure of the Hadamard transform and the matrix inversion for reconstruction suggest the potential for real-time image processing.
Key Results
- High-Fidelity Reconstruction: The reconstructed images from both full and sub-sampled Hadamard acquisitions show remarkable agreement with the simulated target, demonstrating the feasibility of the technique.
- Significant Acquisition Time Reduction: By using only 1/9th of the full Hadamard acquisitions (sampling every 3rd column and row), they achieved a 9x reduction in image acquisition time while maintaining high image quality.
- Subwavelength Resolution Achieved: The system demonstrated the capability of achieving subwavelength resolution (1.24 mm pixel size for 2 mm wavelength) in passive MMW imaging.
Limitations and Future Work
- SNR Degradation: The paper acknowledges a reduction in SNR when using the Hadamard mask, which is a trade-off for improved resolution. Further work could explore methods to mitigate this SNR loss.
- Relaxation Method Efficiency: While the relaxation method is effective, its computational efficiency for very large sub-sampling ratios or complex scenes could be an area for future optimization.
- Passive Imaging Limitations: The current setup uses an incandescent lamp as an illumination source. Future work could focus on improving the performance in truly passive scenarios where the object's own thermal emission is the signal.
- Further Sub-sampling: The paper suggests that if reconstructed image quality is poor, additional sampling cycles can be acquired. Future research could investigate optimal sampling strategies and adaptive acquisition schemes.
Summary updated