Distributed Deblurring of Large Images of Wide Field-Of-View

Rahul Mourya, André Ferrari, Rémi Flamary, Pascal Bianchi, Cédric RichardarXiv 2017 · DOI · arXiv

Abstract

Image deblurring is an economic way to reduce certain degradations (blur and noise) in acquired images. Thus, it has become an essential tool in high-resolution imaging in many applications, such as astronomy, microscopy, and computational photography. In applications such as astronomy and satellite imaging, the size of acquired images can be extremely large (up to gigapixels) covering a wide field-of-view and suffering from shift-variant blur. Most existing image deblurring techniques are designed to work efficiently on centralized computing systems with multiple processors and shared memory. Consequently, the largest image that can be handled is limited by the physical memory available on the system. In this paper, we propose a distributed nonblind image deblurring algorithm in which several connected processing nodes process simultaneously different portions of a large image while maintaining coherency to obtain a single crisp image. We consider certain approximations that trade off between the quality of the deblurred image and the computational resources required. Experimental results show that our algorithm produces image quality similar to existing centralized techniques while allowing distribution, making it cost-effective for extremely large images.

Introduction: The Gigapixel Challenge

In fields like astronomy and satellite imaging, capturing images at gigapixel resolutions is becoming standard. However, these images often suffer from shift-variant blur, where the point-spread function (PSF) changes across the wide field-of-view due to optical aberrations or atmospheric turbulence.

Traditional deblurring techniques are "centralized," requiring the entire image and its blur operators to reside in the shared memory of a single system. For a gigapixel image, this creates a massive memory bottleneck; the physical RAM required simply doesn't exist on standard workstations. The challenge lies in deblurring a massive image in pieces while ensuring the final result is a single, coherent, crisp image without "stitching" artifacts or structural inconsistencies at the boundaries.

Methodology: Divide and Conquer with Consensus

The proposed algorithm reformulates the standard Maximum-A-Posteriori (MAP) estimation into a distributed constrained minimization problem solved via the Douglas-Rachford (D-R) splitting algorithm.

  1. Image Partitioning: The large image is split into overlapping blocks, each assigned to a separate processing node.
  2. Local Solvers: Each node uses a quasi-Newton method (VMLM-B) to handle the smooth but non-linear optimization of the data-fidelity and regularization terms, specifically using a Huber loss regularizer.
  3. The Consensus Step: This is the "secret sauce." Between iterations of local deblurring, nodes exchange information about their overlapping pixels. They perform a weighted averaging (consensus) based on 2D first-order interpolation weights. This ensures that adjacent blocks "agree" on what the underlying crisp image should look like, effectively smoothing out the transitions and the varying blur.
Weighting functions and spatial overlaps for shift-invariant and shift-variant blur cases.

Distributed Architecture

The system is truly distributed and synchronous, implemented using the Julia programming language and the Message-Passing-Interface (MPI). A key advantage of this framework is that it requires no master node. Nodes only need to communicate with their immediate neighbors that share overlapping blocks. This keeps communication overhead manageable, as the local solver steps are significantly more computation-intensive than the network exchange.

Distributed communication topology showing neighbor-to-neighbor information exchange between processing nodes.

Experimental Results and Discussion

The researchers evaluated the method using standard benchmark images (Lena, Barbara, Pentagon) with simulated PSFs, including Airy disks and Gaussian kernels with varying widths to mimic optical coma.

  • Quality Parity: The distributed algorithm produced Signal-to-Noise Ratio (SNR) and Structural Similarity Index (SSIM) values nearly identical to centralized methods.
  • Visual Coherency: Unlike "independent" deblurring, which shows clear block boundaries, the proposed method produced seamless images with no transition artifacts.
  • Robustness: The method proved effective even when the density of sampled PSFs was relatively low, showing that the consensus mechanism is robust at approximating smooth blur transitions.
  • Efficiency: The local solver step takes roughly 600 to 800 times longer than the communication step, making the algorithm ideal for distributed clusters.

Conclusion

While centralized deblurring is faster for small to moderate images due to iterative overhead, this distributed approach is the only viable path for "extremely large" images that exceed the memory capacity of a single machine. Future work will focus on extending this framework to distributed blind deblurring, where processing nodes must also distributively estimate the blur kernels across the field-of-view.

Summary updated

Abstract

Image deblurring is an economic way to reduce certain degradations (blur and noise) in acquired images. Thus, it has become essential tool in high resolution imaging in many applications, e.g., astronomy, microscopy or computational photography. In applications such as astronomy and satellite imaging, the size of acquired images can be extremely large (up to gigapixels) covering wide field-of-view suffering from shift-variant blur. Most of the existing image deblurring techniques are designed and implemented to work efficiently on centralized computing system having multiple processors and a shared memory. Thus, the largest image that can be handle is limited by the size of the physical memory available on the system. In this paper, we propose a distributed nonblind image deblurring algorithm in which several connected processing nodes (with reasonable computational resources) process simultaneously different portions of a large image while maintaining certain coherency among them to finally obtain a single crisp image. Unlike the existing centralized techniques, image deblurring in distributed fashion raises several issues. To tackle these issues, we consider certain approximations that trade-offs between the quality of deblurred image and the computational resources required to achieve it. The experimental results show that our algorithm produces the similar quality of images as the existing centralized techniques while allowing distribution, and thus being cost effective for extremely large images.
BibTeX
@misc{mourya2017distributed,
  author = {Rahul Mourya and André Ferrari and Rémi Flamary and Pascal Bianchi and Cédric Richard},
  title = {{Distributed Deblurring of Large Images of Wide Field-Of-View}},
  year = {2017},
  eprint = {1705.06603},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1705.06603},
}
Figure