💻 computer science

Shift Variant Image Degradation and Restoration Using Singular Value Decomposition

This paper proposes a singular value decomposition (SVD)-based framework for restoring images degraded by shift-variant motion blur, utilizing a singular-value energy retention criterion to systematically select singular values and effectively recover image details while controlling noise amplification across various motion models.

Original authors: Arun D. Kulkarni

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arun D. Kulkarni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a clear photo of a busy street, but your camera is shaking, or the objects you are filming are moving at different speeds. The result is a blurry mess where some parts of the image are slightly fuzzy, while others are smeared into long streaks. This is what scientists call shift-variant image degradation.

In the world of photography and imaging, "shift-invariant" blur is like a uniform fog covering the whole picture equally. You can fix that with a standard recipe. But "shift-variant" blur is like a fog that gets thicker on the left side of the photo and thinner on the right, or a smear that changes shape as you move across the image. Fixing this is much harder because there is no single recipe to apply to the whole picture.

Here is how the author, Arun D. Kulkarni, proposes to solve this puzzle using a mathematical tool called Singular Value Decomposition (SVD).

The Problem: The "Noisy" Puzzle

Think of a degraded image as a giant, broken puzzle. The "blur" is the process that broke the pieces, and "noise" (like static on an old TV) is the dust that got mixed in.

To fix the image, you need to reverse the breaking process. However, because the blur changes across the image, the "breaking machine" is incredibly complex and unstable. If you try to simply reverse it (like rewinding a tape), the tiny bits of dust (noise) get amplified massively, turning your restored image into a grainy, unrecognizable mess. It's like trying to un-bake a cake; if you aren't careful, you don't just get the cake back, you also get a cloud of flour everywhere.

The Solution: The "Energy Filter"

The paper proposes a smart way to reverse this process using SVD. You can think of SVD as a way to take that complex "breaking machine" and break it down into a list of ingredients, ranked from the most important to the least important.

The Ingredients (Singular Values): Imagine the image restoration process is a recipe with 128 ingredients. The first few ingredients are the "big flavors" (the main shapes and details of the image). The last few ingredients are "tiny specks" (fine details mixed with a lot of noise).
The Danger: If you try to use all 128 ingredients to rebuild the image, the tiny specks at the end will overwhelm the recipe, making the result taste terrible (full of noise).
The Old Way (Truncation): Previous methods tried to fix this by simply throwing away the last 50 ingredients. But this is risky. If you throw away too many, you lose important details (like the texture of a face). If you throw away too few, the noise ruins the picture. It's a guessing game.

The Paper's New Idea:
Instead of just throwing away the "tiny speck" ingredients, the author suggests a 99% Energy Retention Criterion.

The Analogy: Imagine you have a bucket of water (the image energy) and you want to keep 99% of it. You look at your list of ingredients and say, "I will keep every ingredient until I have captured 99% of the total water."
The Twist: For the remaining tiny ingredients (the last 1% that we don't need to keep the water level up), instead of throwing them in the trash, we turn down their volume. We don't delete them; we just make them whisper instead of shout.

This way, we keep the important details safe, but we mute the noisy parts that would otherwise ruin the picture.

The Experiments: Three Types of Blurs

To test this "volume control" method, the author created three specific types of "shaking cameras" (mathematical models) and tried to fix images taken with them:

Bi-directional Linear Motion: Imagine a camera moving back and forth in a straight line, but the speed changes depending on where you look in the photo.
Gaussian Motion: Imagine the camera shaking in a way that follows a bell-curve pattern (most shaking in the middle, less on the edges), but the "width" of that shake changes across the image.
Simple Harmonic Motion (SHM): Imagine the camera swinging like a pendulum. It spends more time at the ends of the swing and moves fast through the middle. This creates a specific type of blur that is heavier at the edges of the motion.

The author tested this method on three famous test images (a woman named Lena, a crow, and a text image). In every case, the method successfully took the blurry, noisy mess and turned it back into a clear picture, recovering details that other methods missed.

The Bottom Line

This paper presents a new, systematic way to fix blurry photos where the blur changes from one spot to another. Instead of guessing how much of the image to throw away to fix the noise, the author uses a "99% energy rule" to decide exactly how much to keep and how much to turn down. It's like having a precise volume knob for the noise, ensuring the music (the image) comes through clearly without the static.

The author notes that this method works well for these specific types of motion blurs and suggests it could be expanded in the future to handle even more complex 2D blurs or combined with modern AI techniques, but the current work focuses strictly on proving this mathematical "volume control" works for the three motion types tested.

Technical Summary: Shift Variant Image Degradation and Restoration Using Singular Value Decomposition

Problem Statement
Image degradation in practical imaging systems often arises from factors such as camera or object motion, optical aberrations, atmospheric turbulence, and sensor imperfections. While many restoration techniques assume a shift-invariant Point Spread Function (PSF)—allowing the degradation to be modeled as a simple convolution—real-world scenarios frequently exhibit shift-variant degradation. In these cases, the PSF varies across the image field, rendering conventional Fourier-domain deconvolution methods inapplicable. Shift-variant restoration constitutes a large-scale, ill-posed inverse problem where the degradation operator is represented by a non-Toeplitz matrix. Direct inversion of such matrices is unstable because small singular values amplify measurement noise and numerical errors, leading to unacceptable restoration artifacts. Existing regularization methods, such as Tikhonov regularization and Truncated Singular Value Decomposition (TSVD), face challenges in parameter selection; TSVD, in particular, risks discarding useful image information if the truncation threshold is too aggressive, or failing to suppress noise if it is too lenient.

Methodology
The paper proposes a modified Singular Value Decomposition (SVD)-based framework for restoring images degraded by shift-variant motion blur. The methodology proceeds as follows:

Mathematical Formulation: The degradation process is modeled as a linear system $g = Hf + \epsilon$ , where $g$ is the observed image, $f$ is the original image, $\epsilon$ is additive noise, and $H$ is a shift-variant degradation matrix. Unlike shift-invariant systems, the rows of $H$ correspond to position-dependent PSFs rather than a single kernel.
SVD Analysis: The degradation matrix $H$ is decomposed into singular vectors and singular values ( $\omega_i$ ). The solution to the inverse problem is expressed as a summation of components weighted by the inverse of the singular values. The paper notes that components associated with small singular values are highly sensitive to noise.
Energy-Retention Criterion: To address the instability of direct inversion without arbitrarily discarding information (as in standard TSVD), the authors introduce a singular-value energy retention criterion. The number of singular values to retain is determined by a cumulative energy threshold. Specifically, the method selects a cutoff $K$ such that 99% of the total singular-value energy is preserved.
Attenuation Strategy: Instead of eliminating the components corresponding to the $K$ small singular values, the proposed method attenuates them using a scaling factor derived from the singular-value magnitudes. This approach aims to reduce noise amplification while preserving useful image information that might be lost through hard truncation.
PSF Models: The framework is tested using three representative one-dimensional shift-variant motion PSFs:
- Bidirectional Linear Motion: Blur length varies linearly with position.
- Gaussian Motion: Blur width (standard deviation) varies with position.
- Simple Harmonic Motion (SHM): Amplitude of oscillation varies with position, reflecting the probability density of harmonic displacement.

Key Contributions
The paper identifies two primary contributions:

Modified SVD Restoration Method: The development of a restoration algorithm that utilizes a singular-value energy retention criterion (specifically 99% energy) to systematically select effective singular components. This method attenuates, rather than eliminates, the contribution of small singular values to balance noise suppression and detail preservation.
Algorithmic Framework for Shift-Variant Blur: The formulation of an algorithm to construct shift-variant motion blur kernels and apply the proposed SVD-based restoration to images degraded by these specific non-uniform models.

Experimental Results
The proposed method was evaluated using three standard test images (Lena, Crow, and Text), each sized 128x128 pixels. The images were degraded using the three shift-variant PSFs described above, with additive random Gaussian noise introduced to simulate a Signal-to-Noise Ratio (SNR) of approximately 50.07 dB.

Bi-Directional Linear PSF: The degradation matrix exhibited a condition number of $1.2849 \times 10^{17}$ . Under the 99% energy criterion, 64 small singular values were attenuated, while 64 large singular values were retained.
Gaussian PSF: The condition number was $3.442 \times 10^8$ . The 99% energy criterion resulted in retaining 23 large singular values and attenuating 105 small ones.
Simple Harmonic Motion (SHM) PSF: The condition number was $1.4752 \times 10^4$ . The criterion led to retaining 86 large singular values and attenuating 42 small ones.

Experimental results demonstrated that the proposed approach effectively recovered image details and reduced blur artifacts across all three motion models. The varying number of retained singular values for each model highlighted the necessity of an adaptive threshold based on the specific degradation characteristics.

Significance and Claims
The paper claims that the proposed framework offers a systematic and computationally effective approach for restoring images affected by shift-variant motion blur. By relying on an energy-retention criterion, the method provides a physically meaningful mechanism for parameter selection, avoiding the sensitivity issues associated with conventional truncation thresholds. The authors assert that the singular-value representation offers insight into the ill-conditioned nature of the restoration problem and enables the development of stable inversion techniques. The work concludes that this SVD-based framework is suitable for shift-variant problems and suggests future extensions to two-dimensional spatially varying PSFs and integration with deep-learning methods.

The Problem: The "Noisy" Puzzle

The Solution: The "Energy Filter"

The Experiments: Three Types of Blurs

The Bottom Line

More like this