Shift Variant Image Degradation and Restoration Using Singular Value Decomposition
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Shift Variant Image Degradation and Restoration Using Singular Value Decomposition".
Jane: Here's the gist of the paper. Shift-variant image degradation is frequently encountered in practical imaging systems where the point spread function (PSF) varies across the image field due to motion, optical aberrations, atmospheric turbulence, or sensor-related effects This paper proposes a singular value decomposition (SVD)-based framework for restoring images degraded by shift-variant motion blur The proposed SVD-based restoration algorithm is applied to three degraded images and demonstrates its effectiveness in recovering image details and reducing blur artifacts under different motion models Problem Formulation The degradation process is modeled as a linear system whose structure reflects the position-dependent blur characteristics The shift variant imaging system is represented by an integral equation involving the shift variant point spread function, f(x, y), and additive random noise, ε(x, y) For a discrete image consisting of N pixels, the degradation matrix H is constructed where each row corresponds to a PSF associated with a particular observation location The complete degradation matrix is formed by placing the location-dependent PSFs as successive rows, producing a full system matrix that accurately models spatially varying image blur SVD Analysis and Criterion SVD provides a powerful mathematical framework for analyzing and solving illconditioned inverse problems by decomposing the degradation operator into orthogonal singular vectors and corresponding singular values The solution is extremely sensitive to the β-components associated with the small singular values of the kernel matrix To determine an appropriate value of K, a 99% energy-retention criterion is employed, selecting K so that 99% of the total singular-value energy is retained The proposed attenuation factor is applied to β-components corresponding to small singular values, which reduces the noise amplification due to small singular values Degradation Models Considered Three representative shift-variant motion point spread functions (PSFs) are considered for evaluation: 1.
Tom: Bidirectional linear motion blur, where the blur length varies with position.
Tom: Gaussian motion blur, where the standard deviation varies across the image through a position-dependent standard deviation σ(x).
Paper summary: Tom: So wrapping up this paper, we're talking about "Shift Variant Image Degradation and Restoration Using Singular Value Decomposition". The authors are proposing this SVD framework with their energy retention criterion to handle those tricky shift variant blurs.
Jane: What does that mean for the average person listening right now who might be driving or walking and seeing a blurry photo? It means we have a better tool than just blindly applying standard truncation methods to try and clean up those images.
Meng: It's about having a systematic control mechanism over noise amplification, which is crucial when you're trying to deploy restoration models that need to be robust under varying conditions.
Lu: The paper suggests that by defining the number of singular values based on a percentage of cumulative singular value energy, they stabilize the inverse solutions, which ties back into the theoretical basis for truncated singular value decomposition and Tikhonov regularization discussed in earlier literature <ref:2606.25818#pg3>.
Lalam: I'm sure it's quite confident in its answer because it was tested against three distinct types of spatially variant blur, showing effectiveness across those different motion models.
Tom: The title and the authors are Arun D. Kulkarni and colleagues, and they are showing that a specific SVD modification provides a more stable way to handle these position-dependent point spread functions compared to what was previously done with standard TSVD.
Jane: So in simple terms, this work gives us a mathematical recipe for choosing the right components from the singular value decomposition to fix images that have motion blur that changes depending on where you are looking.
Meng: It provides a way to reduce noise amplification while preserving useful image information through an energy-based selection process, which is something we need when deploying these restoration techniques at scale.
Lu: It's a strong theoretical step toward making the parameter selection less sensitive and more physically motivated for solving these ill-posed problems eight nine <ref:2606.25818#pg3>.
Lalam: The paper shows that their approach effectively restores blurred images and recovers structural details under those different degradation conditions, demonstrating its practical potential.
Conclusion: Tom: So, we're wrapping up this look at "Shift Variant Image Degradation and Restoration Using Singular Value Decomposition." Basically, these authors are proposing a specific mathematical method to fix images where the blur pattern changes depending on where you look in the picture.
Jane: It sounds complicated, but what they’re actually doing is using singular value decomposition, or SVD, which is just a way to break down how much information and noise is packed into an image.
Lu: They construct a degradation matrix that has different blur patterns for every single spot on the image, and then they use SVD to analyze that whole mess. It lets them separate the clean image from all those blurry artifacts at once.
Meng: So, if I'm running this on a real system, what's the actual mechanism they use to decide which parts of that decomposition are good and which ones are just noise? I need to know how they select those components.
Lalam: They use an energy retention criterion. They calculate the total energy from all the singular values, and they keep only enough singular values so that ninety-nine percent of that original image information is still there after the process.
Tom: That’s a concrete number, which is good because it stops them from just picking a fixed number of components and hoping for the best. It makes their selection process systematic instead of arbitrary.
Jane: It means they are not guessing how much noise to remove; they're using the actual energy content of the image to guide their choice between keeping signal and cutting out noise amplification caused by those small singular values.
Lu: That’s the core idea, I think. They apply a factor to those components associated with the very small singular values to reduce how much noise they amplify during the fix. It’s a regularization technique built directly into the SVD structure of this specific problem.
Meng: If it works like that, does it mean we can deploy this kind of restoration system reliably in production where things are constantly moving and changing? Or is it still too sensitive to the exact noise characteristics we don't know for sure?
Lalam: The experimental results show that this energy-based strategy provides a very stable way to select the effective singular components, which improves the stability of their restoration significantly. It’s much better than just using conventional truncation methods where you just cut off the tail.
Tom: It sounds like they managed to solve a problem where the inverse math is really tough because it's so ill-conditioned—meaning tiny changes in input cause huge changes in output. They’re taming that sensitivity with a physical constraint, which is what we need for engineering.
Jane: So the big picture here is that this SVD framework offers a systematic and physically meaningful way to handle those spatially variant blur problems, giving us a tool that preserves detail while actively fighting noise amplification across different motion models.
Lu: It moves the problem from being just an inverse problem to one where you have an established criterion for selecting the right mathematical components based on data retention. That’s a pretty solid theoretical improvement for solving these kinds of ill-posed issues.
Meng: I'm interested in seeing how this translates when we run it against something messy, like real-world sensor noise that isn't just simple Gaussian noise, but something more complex. We need to know if this energy retention holds up under heavy stress.
Lalam: The paper tested it on three different motion models—bidirectional linear blur, Gaussian blur with variable standard deviation, and simple harmonic motion—and it worked well across all of them. It’s shown to be effective in recovering structural details under those various conditions.
Tom: So the authors are showing that this single SVD-based framework isn't just a toy model; it handles a variety of physical blur behaviors without needing a completely custom solution for every single motion type.
Jane: It really boils down to having a robust, energy-aware selection mechanism built into the SVD process itself, which makes the restoration much more stable and less prone to those nasty noise exaggerations we see in standard methods.
Lu: This approach provides an analytical advantage by coupling the power of SVD decomposition with an explicit regularization mechanism that is tied directly to how much signal energy you want to keep. That’s a solid direction for solving these kinds of spatially variant image restoration problems.
Arun D. Kulkarni
University of Texas at Tyler
cs.CV
Submitted: 2026-06-24
Updated: 2026-11-15
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 79/100
The gist: Shift-variant image degradation is frequently encountered in practical imaging systems where the point spread function (PSF) varies across the image field due to motion…
Key concepts
- Shift-Variant Point Spread Function (PSF)
- This describes how the blur pattern changes across an image. Instead of a single, fixed blur shape, the PSF varies depending on where you are looking in the picture. This variation can be caused by things like camera movement or optical imperfections during capture.
- Singular Value Decomposition (SVD)
- SVD is a mathematical tool that breaks down a complex degradation operator into orthogonal components. It separates the system's effects into singular vectors and their corresponding singular values. This decomposition helps identify which parts of the data contribute most to the image structure versus noise.
- Energy-Retention Criterion
- This is a method used to decide how many components from the SVD should be kept for restoration. The criterion selects components that retain 99% of the total singular value energy. This ensures that important image information is preserved while aggressively filtering out noise associated with very small singular values.
Terminology
Summary
Shift-variant image degradation is frequently encountered in practical imaging systems where the point spread function (PSF) varies across the image field due to motion, optical aberrations, atmospheric turbulence, or sensor-related effects This paper proposes a singular value decomposition (SVD)-based framework for restoring images degraded by shift-variant motion blur The proposed SVD-based restoration algorithm is applied to three degraded images and demonstrates its effectiveness in recovering image details and reducing blur artifacts under different motion models
Problem Formulation
The degradation process is modeled as a linear system whose structure reflects the position-dependent blur characteristics The shift variant imaging system is represented by an integral equation involving the shift variant point spread function, f(x, y), and additive random noise, ε(x, y) For a discrete image consisting of N pixels, the degradation matrix H is constructed where each row corresponds to a PSF associated with a particular observation location The complete degradation matrix is formed by placing the location-dependent PSFs as successive rows, producing a full system matrix that accurately models spatially varying image blur
SVD Analysis and Criterion
SVD provides a powerful mathematical framework for analyzing and solving illconditioned inverse problems by decomposing the degradation operator into orthogonal singular vectors and corresponding singular values The solution is extremely sensitive to the β-components associated with the small singular values of the kernel matrix To determine an appropriate value of K, a 99% energy-retention criterion is employed, selecting K so that 99% of the total singular-value energy is retained The proposed attenuation factor is applied to β-components corresponding to small singular values, which reduces the noise amplification due to small singular values
Degradation Models Considered
Three representative shift-variant motion point spread functions (PSFs) are considered for evaluation:
-
Bidirectional linear motion blur, where the blur length varies with position. p. 4
Where the paper says it
Point Spread Functions for Shift Variant Motion Blur A bidirectional linear motion blur occurs when the object moves along a straight line during the exposure interval. If the blur length varies with position, the PSF becomes shiftvariant.
-
Gaussian motion blur, where the standard deviation varies across the image through a position-dependent standard deviation σ(x). p. 4
Where the paper says it
Shift variant Gaussian PSF Gaussian motion blur models situations where the probability of displacement follows a Gaussian distribution. The blur spread varies across the image through the position-dependent standard deviation 𝜎(𝑥).
-
Simple harmonic motion (SHM) blur, where the amplitude varies spatially as A(x). p. 6
Where the paper says it
Degraded and Restored Images for Bi-Directional Linear PSF C. Shift-Variant Simple Harmonic Motion PSF Unlike linear and Gaussian motion PSFs, the SHM PSF exhibits higher weighting near the extreme motion positions, reflecting the longer residence time of the oscillating object at kernel size is 128×128, and the amplitude in (21) varies 2 to 4.
Experimental Findings
Experimental results obtained using three test images demonstrated that the proposed approach effectively restored blurred images and recovered important structural details under different degradation conditions The energy-retention-based strategy provided a systematic and physically meaningful mechanism for selecting the effective singular components, thereby improving restoration stability Overall, the proposed framework combines the analytical advantages of SVD with an energy-based regularization mechanism, providing an effective and computationally efficient approach for shift-variant image restoration The method avoids the sensitivity associated with conventional truncation methods by attenuating noise amplification while preserving useful image information
ACKNOWLEDGEMENT
The author would like to express sincere gratitude to the late Prof. S. C. Sahasrabudhe for his invaluable guidance, encouragement, and support for this work. p. 8 ACKNOWLEDGEMENT
The author would like to express sincere gratitude to the late
Prof. Sahasrabudhe for his invaluable guidance,
encouragement, and support for this work.Where the paper says it
REFERENCES
[1] R. C. Gonzalez and R. p. 1 Classical restoration techniques are
primarily based on inverse filtering, Wiener filtering, and
regularized least-squares formulations. Comprehensive
treatments of these methods are provided by Gonzalez and
Woods [1]. Gonzalez and R. Woods, Digital Image Processing, 4th ed.Where the paper says it
Where the paper says it
[2] Y. P. Guo and R. K. Ward, “Blind Restoration of Images Degraded by Space-Variant Blurs Using Iterative Algorithms,” Image and Vision Computing, vol. p. 8 Pearson, 2018. Ward, “Blind Restoration of Images Degraded by
Space-Variant Blurs Using Iterative Algorithms,” Image and Vision
Computing, vol.Where the paper says it
[3] J. G. Nagy and D. p. 1 Their
work demonstrated that image and blur parameters could be
estimated simultaneously, even when complete PSF
information was unavailable. Nagy and O'Leary [3]
subsequently formulated spatially variant image restoration as
a large-scale linear inverse problem and developed matrixbased computational techniques for handling non-uniform
blur.Where the paper says it
[4] S. B. Hadj, L. B.-Féraud, and G. p. 8 1063-1082, July 1998. -Féraud, and G. -Féraud, and G. Aubert, “Space Variant Blind Image
Restoration,” SIAM Journal on Imaging Sciences, vol.Where the paper says it
Where the paper says it
[5] Y. Zhang, X. Wang, and C. p. 2 [18] proposed
ASANet to restore images affected by wide-field aberrations. Wang et al. Wang, and C. Wang, “Restoration of Space-Variant
Blurred Image Based on Motion-Blurred Target Segmentation,” Journal
of Systems Engineering and Electronics, vol. Wang, “Restoration of Space-Variant
Blurred Image Based on Motion-Blurred Target Segmentation,” Journal
of Systems Engineering and Electronics, vol. 191–196, Apr.Where the paper says it
Where the paper says it
Where the paper says it
[6] H. Cheong et al. p. 1 [5] introduced a restoration method based on motionblurred target segmentation, where locally blurred regions are
identified and restored independently to improve
reconstruction quality. Cheong et al.Where the paper says it
[7] A. N. Tikhonov and V. p. 0 Numerous approaches based on
numerical inversion have been proposed; however, direct
inversion often produces unstable solutions because
measurement errors, numerical round-off errors, and noise are
significantly amplified by the small singular values of the
degradation matrix. Although several regularization
techniques have been developed for image restoration,
methods such as Tikhonov regularization, Wiener filtering,
and iterative restoration approaches have certain limitations
when applied to shift-variant degradation models. Tikhonov and V. Arsenin, Solutions of Ill-Posed Problems.Where the paper says it
Where the paper says it
[8] P. C. Hansen, Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion. p. 8 New
York, NY, USA: Wiley, 1977. Hansen, Rank-Deficient and Discrete Ill-Posed Problems: Numerical
Aspects of Linear Inversion.Where the paper says it
[9] P. C. Hansen, Discrete Inverse Problems: Insight and Algorithms. p. 8 Philadelphia, PA, USA: SIAM, 1998. Hansen, Discrete Inverse Problems: Insight and Algorithms.Where the paper says it
[10] G. Huang, Y. Liu, and F. p. 8 Philadelphia, PA, USA: SIAM, 2010. Liu, and F. Liu, and F. Yin, “Tikhonov Regularization with MTRSVD
Method for Solving Large-Scale Discrete Ill-Posed Problems,” Journal of
Computational and Applied Mathematics, vol.Where the paper says it
Where the paper says it
[11] H. C. Andrews and B. p. 2 Singular value
decomposition (SVD) has long been recognized as a powerful
mathematical tool for analyzing degradation operators because
it reveals the spectral properties and conditioning of inverse
problems. Andrews and Hunt [11] emphasized the importance
of matrix formulations in image restoration and demonstrated
the usefulness of SVD in inverse filtering. Andrews and B. Hunt, Digital Image Restoration.Where the paper says it
Where the paper says it
[12] A. Rajwade, A. p. 2 SVD provides an
orthogonal decomposition of the degradation matrix into
singular vectors and singular values, allowing the contribution
of different frequency components to be analyzed separately. Rajwade et al. Rajwade, A. Rangarajan, and A. Rangarajan, and A. Banerjee, “Image Denoising Using
the Higher-Order Singular Value Decomposition,” IEEE Trans. Banerjee, “Image Denoising Using
the Higher-Order Singular Value Decomposition,” IEEE Trans. Pattern
Anal. Banerjee, “Image Denoising Using
the Higher-Order Singular Value Decomposition,” IEEE Trans. Pattern
Anal.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[13] M. Elad and M. p. 2 Sparse representation and variational regularization
methods have also played a significant role in image
restoration. Elad and Aharon [13] introduced the K-SVD
algorithm, which learns overcomplete dictionaries for sparse
image representation and denoising. Elad and M. Aharon, “Image Denoising via Sparse and Redundant
Representations Over Learned Dictionaries,” IEEE Trans. Kulkarni
Abstract— Shift-variant image degradation is frequently
encountered in practical imaging systems where the point spread
function (PSF) varies across the image field due to motion,
optical aberrations, atmospheric turbulence, or sensor-related
effects. Unlike shift-invariant, shift-variant degradation presents
significant challenges for image restoration because the
degradation process cannot be represented by a single
convolution kernel. Kulkarni
Abstract— Shift-variant image degradation is frequently
encountered in practical imaging systems where the point spread
function (PSF) varies across the image field due to motion,
optical aberrations, atmospheric turbulence, or sensor-related
effects. Unlike shift-invariant, shift-variant degradation presents
significant challenges for image restoration because the
degradation process cannot be represented by a single
convolution kernel.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[14] L. Rudin, S. Osher, and E. p. 2 Their approach
demonstrated that sparse signal models can effectively
preserve important image structures while suppressing noise. Rudin, Osher, and Fatemi [14] proposed total variation (TV)
regularization, which became one of the most influential edgepreserving restoration techniques. Osher, and E. Fatemi, “Nonlinear Total Variation Based
Noise Removal Algorithms,” Physica D, vol.Where the paper says it
Where the paper says it
[15] C. Dong, C. Loy, K. He, and X. Tang, “Image Super-Resolution Using Deep Convolutional Networks,” IEEE Trans. p. 8 259–268, 1992. Tang, “Image Super-Resolution Using
Deep Convolutional Networks,” IEEE Trans.Where the paper says it
[16] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. p. 8 295–307, 2016. Meng, and L. Meng, and L. Zhang, “Beyond a
Gaussian Denoiser: Residual Learning of Deep CNN for Image
Denoising,” IEEE Trans.Where the paper says it
Where the paper says it
[17] O. Ronneberger, P. p. 2 [16] introduced DnCNN, which employs residual learning for
image denoising and significantly improves restoration
accuracy. Ronneberger et al. Ronneberger, P. Fischer, and T.Where the paper says it
Where the paper says it
[18] Y. Tian et al. p. 2 [17] proposed U-Net, whose
encoder-decoder structure and skip connections have inspired
numerous restoration networks. For spatially varying blur, Tian et al. Tian et al. , “Restoration of Spatially Variant Blurred Images with
Wide-Field Aberrations Using ASANet,” Sensors, vol.Where the paper says it
Where the paper says it
[19] J. Wang, J. Liu, and X. p. 8 Philadelphia, PA, USA: SIAM, 2010. Liu, and F.Where the paper says it
[20] J. Zhang and F. p. 1 [4] proposed a variational framework for space-variant blind
image restoration in which both the latent image and the
spatially varying PSF are estimated simultaneously. Zhang et
al.Where the paper says it
[21] P. Jia et al., “Image Restoration with Point-Spread Function Regularization and Active Learning,” Monthly Notices of the Royal Astronomical Society, vol. p. 8 0075
[21] P. , “Image Restoration with Point-Spread Function
Regularization and Active Learning,” Monthly Notices of the Royal
Astronomical Society, vol. , “Image Restoration with Point-Spread Function
Regularization and Active Learning,” Monthly Notices of the Royal
Astronomical Society, vol. 6581–6590, Jan.Where the paper says it
Where the paper says it
[22] J. Li, H. Wang, Y. Li, and H.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can restore images degraded by shift-variant motion blur by using a modified Singular Value Decomposition framework that controls noise amplification through an energy-retention criterion. This method selects the number of retained singular values based on a
99% energy-retention criterion,
providing a systematic and physically meaningful selection mechanism compared to conventional Truncated Singular Value Decomposition (TSVD). -
Improved AI system can generate shift-variant motion blur kernels by modeling different spatial variations such as bidirectional linear motion, Gaussian motion, and simple harmonic motion. These specific point spread functions are used to create the
degradation matrices
which accurately model spatially varying image blur across different locations. -
Improved AI system can perform stable image restoration by attenuating the influence of small singular values using an attenuation factor derived from the energy-retention criterion. This scaling factor is applied to components associated with small singular values, reducing noise amplification while preserving useful image information, which offers a more accurate estimate than conventional truncation methods.
-
Improved AI system can be extended to complex degradation scenarios by incorporating
adaptive and data-driven parameter selection strategies.
This future direction will involve integrating the proposed SVD framework with iterative and deep learning-based restoration methods for more challenging imaging problems.
Abstract
Shift-variant image degradation is frequently encountered in practical imaging systems where the point spread function (PSF) varies across the image field due to motion, optical aberrations, atmospheric turbulence, or sensor-related effects. Unlike shift-invariant, shift-variant degradation presents significant challenges for image restoration because the degradation process cannot be represented by a single convolution kernel. This paper proposes a singular value decomposition (SVD)-based framework for restoring images degraded by shift-variant motion blur. The proposed approach determines the contribution of small singular values using a singular-value energy retention criterion. Specifically, the number of small singular values is selected based on a specified percentage of cumulative singular-value energy, providing a systematic approach for controlling noise amplification while preserving useful image information. The degradation model is formulated using a position-dependent PSF represented by a shift-variant imaging operator. Three representative one dimensional shift-variant motion PSFs are considered: bidirectional linear motion, Gaussian motion, and simple harmonic motion. The image degradation process is modeled as a linear system, and SVD is employed to analyze and invert the corresponding degradation operator. The singular-value representation provides insight into the ill-conditioned nature of the restoration problem and enables the development of stable inversion techniques. The proposed SVD-based restoration algorithm is applied to three degraded images. Experimental results demonstrate the effectiveness of the proposed approach in recovering image details and reducing blur artifacts under different motion models.
Sources
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs