Principal Component Analysis-Based Terahertz Self-Supervised Denoising and Deblurring Deep Neural Networks

This paper proposes a principal component analysis-based self-supervised deep neural network (THz-SSDD) that effectively addresses the simultaneous challenges of low-frequency blurring and high-frequency noise in terahertz amplitude images by leveraging a Recorrupted-to-Recorrupted learning strategy and PCA reconstruction without requiring labeled data.

The Gist

The Problem: The "Fuzzy and Grainy" Terahertz Camera

Imagine you have a special camera that can see inside objects (like wood, plastic, or composite materials) without breaking them. This is a Terahertz (THz) camera. It's like having X-ray vision for industrial quality control.

However, this camera has a major flaw. The images it takes are like a photo taken on a very foggy day with a shaky hand:

1. The Blur: The edges of things look soft and fuzzy (like looking through a dirty window). This happens because the camera's "lens" isn't perfect.
2. The Grain: The image is covered in static or "snow" (like an old TV channel with no signal). This is random noise that hides the details.

The Dilemma:
Usually, if you try to fix the blur, you make the grain worse. If you try to clean up the grain, you make the blur worse. It's like trying to sharpen a pencil while simultaneously sanding off the tip. Scientists usually had to pick one problem to fix, or they had to manually tweak the settings for every single image, which is slow and frustrating.

The Solution: The "Smart Photo Editor" (THz-SSDD)

The authors of this paper built a new AI tool called THz-SSDD. Think of it as a super-smart photo editor that can fix both the blur and the grain at the same time, without needing a human to tell it how.

Here is how it works, broken down into three simple steps:

1. The "Deconstruct and Rebuild" Strategy (PCA)

Imagine you have a messy pile of 100 different colored Lego bricks representing your image. Some bricks are the important shape (the object), and some are just random trash (noise).

• The Trick: Instead of trying to clean the whole pile at once, the AI sorts the bricks into groups based on how much they look like the "main shape."
• The Result: It keeps the top 5 groups of bricks (the most important parts of the image) and throws away the rest of the messy pile. This simplifies the problem, making it easier to clean.

2. The "Practice Makes Perfect" Strategy (Self-Supervised Learning)

Usually, to teach an AI to clean a photo, you need to show it a "dirty" photo and the "perfect" clean version of that same photo. But in real life, we rarely have the "perfect" version of a terahertz image.

• The Analogy: Imagine trying to learn how to clean a muddy window, but you don't have a clean window to compare it to.
• The Solution: The AI uses a strategy called "Recorrupted-to-Recorrupted" (R2R). It takes the dirty window, adds more mud to it in a specific way, and then tries to clean it back to the original dirty state. By playing this game of "add mud, then remove mud" over and over, the AI learns exactly what the "mud" (noise) looks like and how to remove it without touching the "glass" (the real image). It learns by doing, not by memorizing answers.

3. The "Reassembly"

Once the AI has cleaned up those top 5 groups of Lego bricks (the principal components), it puts them back together to rebuild the image. Because it cleaned the groups individually before putting them back, the final image is both sharp (no blur) and clear (no grain).

What Did They Test?

To prove this new tool works, they didn't just test it on one thing. They tested it on a variety of "mystery boxes":

• Glass Fiber Plates: To see if it could find hidden holes.
• Burnt Wood: To see if it could spot damage from fire.
• Stretched Plastic: To see if it could find weak spots where the plastic was about to break.
• Mixed Composites: To see if it could find cracks caused by dropping heavy weights on them.

The Result: The AI worked like a charm. It took blurry, noisy images of all these different materials and turned them into crisp, clear pictures. It could find tiny holes and cracks that were previously hidden by the "fog" and "snow."

Why Does This Matter?

In the real world, this means factories can inspect materials faster and more accurately.

• No more manual tweaking: You don't need an expert to adjust settings for every new type of material.
• One tool for all: The same AI works on wood, plastic, and metal composites.
• Safety: It helps ensure that bridges, planes, and buildings made of these materials are safe, because the "hidden" defects are now visible.

In a nutshell: The authors created a smart AI that teaches itself how to clean up bad terahertz images by playing a game of "add noise, then remove noise," and then uses a sorting technique to rebuild the picture perfectly. It's like giving a blurry, grainy photo a magic makeover.

Technical Summary

1. Problem Statement

Terahertz (THz) imaging systems are powerful tools for non-destructive testing (NDT) due to their non-ionizing nature and ability to penetrate various materials. However, THz amplitude images suffer from inherent frequency-dependent degradation:

• High-Frequency Noise: System noise, antenna-induced noise, and thermal fluctuations create significant noise in high-frequency regions.
• Low-Frequency Blurring: Diffraction limits, beam shape, and optical system constraints cause low-frequency blurring, obscuring defect boundaries.

Challenges:

• Simultaneous Processing: Conventional image processing techniques typically address denoising and deblurring separately. Applying them sequentially often leads to suboptimal results or requires manual intervention to determine the processing order.
• Data Scarcity: Supervised deep learning methods require large datasets of paired "clean" and "noisy" images, which are difficult to obtain in THz-TDS (Time-Domain Spectroscopy) because ground-truth clean signals are unavailable.
• Generalization: Existing unsupervised methods often lack robustness when applied to samples with different material properties or measurement modes (e.g., reflection vs. transmission).

2. Methodology

The authors propose a PCA-based THz Self-Supervised Denoising and Deblurring Deep Neural Network (THz-SSDD). The framework integrates Principal Component Analysis (PCA) with a self-supervised learning strategy.

A. Core Architecture: THz-SSDD

• Learning Strategy: The network employs a Recorrupted-to-Recorrupted (R2R) self-supervised learning strategy. Instead of requiring clean ground truth, the model learns to map a noisy image to a differently corrupted version of itself. This exploits the invariance of intrinsic features under repeated corruption.
• Network Structure: The model utilizes two U-Net networks:
  1. One for Denoising.
  2. One for Deconvolution (Deblurring).
• Loss Function: The training objective combines three terms:
  1. Pixel-wise Self-Supervision: Enforces noise removal by comparing the output to a differently corrupted input.
  2. PSF Alignment: Aligns the Point Spread Function (PSF)-convolved output with the target to drive blur inversion.
  3. Regularization: Penalizes spurious high-frequency artifacts to prevent noise amplification.
• PSF Modeling: The Point Spread Function is theoretically derived based on a diffraction-limited Gaussian beam model, allowing the network to learn the specific system blur characteristics.

B. PCA-Based Image Restoration Pipeline

Since single-image processing cannot simultaneously solve low-frequency blurring and high-frequency noise, the authors introduce a PCA-based workflow:

1. Transformation: The 3D THz spectral data cube ($H \times W \times \text{Frequency}$) is reshaped into a 2D matrix where columns represent spectral vectors of pixels.
2. Decomposition: PCA is applied to decompose the data into principal components (PCs).
3. Selection: The top $r$ principal components (specifically the first five in this study) are selected. These components retain 95–99% of the total variance and contain the most significant structural information.
4. Processing: The selected PCs are fed into the trained THz-SSDD network for simultaneous denoising and deblurring.
5. Reconstruction: The processed PCs are reconstructed back into the 3D spectral image cube.

3. Key Contributions

1. Unified Framework: This is the first attempt in THz-TDS to use a single algorithm/network to simultaneously address low-frequency blurring and high-frequency noise, eliminating the need for manual intervention or sequential processing.
2. Self-Supervised Learning: The method operates without labeled clean data. It is trained exclusively on a small set of unlabeled noisy images from a calibration plate (GFRP), making it highly practical for industrial applications where ground truth is unavailable.
3. Generalization: The approach demonstrates robust generalization across diverse materials (wood, polymers, composites) and measurement modes (reflection and transmission), despite being trained on a single sample type.
4. Signal Fidelity Preservation: By combining PCA with self-supervised learning, the method restores image quality while preserving the physical characteristics and amplitude trends of the original THz signals.

4. Experimental Results

The method was evaluated on four distinct sample types:

• Training Sample: Glass Fiber Reinforced Polymer (GFRP) laminate with flat-bottom holes.
• Testing Samples:
  • Pyrolyzed spruce wood (different temperatures).
  • High-Density Polyethylene (HDPE) dog-bone specimens (plastic deformation and necking).
  • Hybrid Basalt/Flax composites (subjected to low-velocity impact damage).

Performance Metrics:

• Visual Quality: The network effectively suppressed Gaussian noise and sharpened blurred defect boundaries. In hybrid composites, it successfully recovered impact damage that was previously obscured by complex fiber patterns and resin inhomogeneity.
• Quantitative Metrics:
  • PSNR (Peak Signal-to-Noise Ratio): Showed significant improvement, particularly in frequency ranges where noise was dominant.
  • RSE (Relative Spectral Error): Used to measure signal deviation. Most samples showed an RSE of approximately 0.4, indicating high signal fidelity. The method successfully restored details without drastically altering the underlying physical signal trends, except in cases of extreme signal attenuation (e.g., heavily damaged composites).

5. Significance

This research addresses a critical bottleneck in THz imaging: the trade-off between noise reduction and resolution enhancement. By leveraging self-supervised learning and PCA decomposition, the proposed THz-SSDD network offers a robust, automated solution for industrial NDT.

• Practicality: It removes the dependency on large, labeled datasets, which are often the limiting factor in deep learning applications for THz.
• Versatility: The ability to generalize from a single training sample to diverse industrial materials (polymers, woods, composites) and different measurement setups makes it a viable candidate for real-world deployment.
• Scientific Impact: The method preserves the quantitative integrity of THz signals, ensuring that the restored images remain valid for extracting physical properties and defect dimensions, not just for visual inspection.