Subsampled terahertz data reconstruction based on spatio-temporal dictionary learning

Vahid Abolghasemi, Hao Shen, Yaochun Shen, Lu Gan

Abstract

Terahertz Pulsed Imaging (TPI) is a powerful non-destructive tool for medical and industrial inspection, but its slow acquisition speed remains a significant challenge. This paper presents a novel framework for reconstructing 3D THz data from highly incomplete samples (as low as 5%). By leveraging spatio-temporal dictionary learning and joint sparse recovery, the proposed method retrieves high-quality structural and spectral information without the need for complex compressive sensing hardware. Experimental results on layered structures and pharmaceutical tablets demonstrate superior performance over traditional wavelet-based methods, achieving high physical measurement accuracy even with 90% missing data.

Introduction

Terahertz Pulsed Imaging (TPI) is a vital tool for non-destructive testing, particularly in medical diagnosis and pharmaceutical inspection. However, standard TPI systems rely on pixel-by-pixel raster scanning, making the acquisition of high-resolution 3D volumes prohibitively slow. While Compressive Sensing (CS) has been proposed to accelerate this process, it often necessitates complex hardware such as specialized spinning disks or Bernoulli masks.

This paper proposes a "software-first" alternative: subsampling. By taking fewer measurements and employing advanced spatio-temporal reconstruction algorithms, the authors aim to achieve the speed benefits of CS without the hardware burden, while maintaining the integrity of both structural and spectral information.

Proposed Methodology

The proposed approach follows a multi-stage pipeline designed to handle noisy and incomplete 3D "datacubes."

Block diagram of the proposed reconstruction approach.
  1. Initial Interpolation: The incomplete data is first roughly filled using Bicubic interpolation.
  2. Spatio-Temporal Block Extraction: The data is partitioned into small 3D blocks ($n_x \times n_y \times b$). This allows the model to capture correlations across both spatial dimensions and the temporal (depth) dimension simultaneously.
  3. Dictionary Learning: Unlike fixed mathematical transforms, this method learns an overcomplete dictionary ($\mathbf{D}$) tailored to the specific dataset using the K-SVD algorithm.
  4. Joint Sparse Recovery: To handle noise, the authors use a Multiple Measurement Vector (MMV) framework with Simultaneous Orthogonal Matching Pursuit (SOMP). This assumes neighboring blocks share similar sparsity patterns, increasing robustness.
  5. Final Reconstruction: The learned dictionary and sparse coefficients are used to solve a convex optimization problem that denoises the data while preserving sharp structural features.

Spatio-Temporal Block Extraction

A key innovation is treating the data as a 3D volume during the learning phase. This exploits the fact that THz pulses change predictably as they penetrate deeper into a material. The algorithm can be tuned by adjusting block sizes ($n_x, n_y, b$) to prioritize either spatial structural features or temporal spectral "fingerprints."

Extraction of 3D spatio-temporal blocks from the datacube Y.

Experimental Setup

The method was evaluated on two distinct THz datasets:

  • T-shape Dataset: A model sample consisting of a T-shaped plastic sheet buried 0.2 mm inside a polythene pellet, used to test the recovery of layered structures.
  • Tablet Data: Low-spatial-resolution data from pharmaceutical tablets, used to evaluate the accuracy of thickness measurements and chemical mapping (spectral analysis).

Results and Discussion

The proposed dictionary learning method consistently outperformed 3D Wavelet transforms and Bicubic interpolation. Key findings include:

  • High Fidelity: For the T-shape data, the method achieved a Signal-to-Noise Ratio (SNR) of 19 dB from only 5% observation.
  • Physical Accuracy: For pharmaceutical tablets, the reconstructed data allowed for thickness and depth measurements with 98.8% and 99.9% accuracy, respectively, even with 90% of the data missing.
  • Chemical Mapping: Analysis using Cosine Correlation Mapping confirmed that the reconstructed spectral information was nearly identical to the original, preserving critical chemical signatures.
  • Denoising Capabilities: By solving training vectors jointly (MMV), the algorithm naturally filters out random noise that does not align with the common underlying structure of the group.

Conclusion

This spatio-temporal dictionary learning framework provides a robust solution for high-speed THz imaging via subsampling. While the method is computationally more intensive than simple interpolation, it offers significant gains in reconstruction quality and physical accuracy. Future work may involve extending the model to the frequency domain to create a joint spatio-temporal-frequency dictionary for even more precise spectral reconstruction in complex chemical mixtures.

Summary updated

BibTeX
@article{abolghasemi2015subsampled,
  author = {Vahid Abolghasemi and Hao Shen and Yaochun Shen and Lu Gan},
  title = {{Subsampled terahertz data reconstruction based on spatio-temporal dictionary learning}},
  journal = {Digit. Signal Process.},
  year = {2015},
  doi = {10.1016/j.dsp.2015.04.010},
}
Figure