Poisson Noise Removal Using Multi-Frame 3D Block Matching

Kireeti Bodduna, Joachim Weickert

Abstract

The 3D block matching (BM3D) filter belongs to the state-of-the-art techniques for eliminating additive white Gaussian noise from single-frame images. There exist four multi-frame extensions of BM3D as of today. In this work, we combine these extensions with a variance stabilising transformation (VST) for eliminating Poisson noise. Our evaluation reveals that the extension which retains the original noise model of the noisy images and additionally has a comprehensive connectivity of 2D and temporal image information at both pixel and patch levels, gives the best results. Additionally, we find a surprising change in performance of one the four extensions due to the specific application of the VST. Finally, we also introduce a simple low-pass filtering as a preprocessing step for the best performing extension. This can give rise to a significant additional improvement of 0.94 dB in the output according to the peak signal to noise ratio.

Introduction

Digital images are frequently corrupted by noise, but the underlying statistical model of that noise dictates the effectiveness of denoising algorithms. While Additive White Gaussian Noise (AWGN) is a common target for research, real-world sensors in fields like medical imaging, astronomy, and electron microscopy are primarily dominated by Poisson noise (or shot noise). Unlike Gaussian noise, Poisson noise is signal-dependent: the variance of the noise increases with the intensity of the signal.

While the 3D Block Matching (BM3D) filter is considered a gold standard for single-frame Gaussian denoising, its application to multi-frame Poisson noise remains a challenge. When multiple shots of the same scene are available, the strategy for combining this information is critical. This paper evaluates four distinct multi-frame strategies to determine the optimal architecture for cleaning up multi-frame Poisson data.

Methodology

To handle the signal-dependent nature of Poisson noise, the authors utilize a Variance Stabilizing Transformation (VST)—specifically the Anscombe transform—to convert Poisson noise into a distribution that approximates Gaussian noise. This allows for the application of Gaussian-based BM3D extensions. Four specific multi-frame strategies were evaluated:

  1. BM3D-1 (Average then Denoise): All noisy frames are averaged into a single image before running standard BM3D.
  2. BM3D-2 (Denoise then Average): Every frame is denoised individually, and the results are then averaged.
  3. BM3D-3 (Single Reference): One frame is selected as a reference, but the algorithm searches for similar patches across all other frames to build the "3D groups" for filtering.
  4. BM3D-4 (Comprehensive Connectivity): Every frame in the stack is treated as a reference frame, allowing the algorithm to search across the entire temporal stack for similar patches.

The "Secret Sauce": BM3D-4$_{\sigma}$

In high-noise environments, the algorithm often struggles to find "similar" patches because the noise masks the underlying structure. To address this "grouping paradox," the authors introduced BM3D-4$_{\sigma}$. This method adds a simple Gaussian low-pass filter as a preprocessing step. This filtered version is used exclusively to help the algorithm find similar patches (grouping), while the actual denoising is performed on the original noisy data.

Experimental Results

The study utilized four classic image processing benchmarks: Lena, House, Bridge, and Peppers. Poisson noise was added with varying "noise peaks" (1 to 5), where lower values indicate higher noise levels.

PSNR (dB) comparison of the four BM3D extensions and the proposed BM3D-4$_{\sigma}$ method across various noise levels and image datasets.

Quantitative evaluation using Peak Signal-to-Noise Ratio (PSNR) confirms the superiority of the BM3D-4 approach. As shown in the performance tables, BM3D-4 consistently outperforms simpler extensions like BM3D-1 and BM3D-2. Interestingly, BM3D-1 (averaging first) performs worse as more frames are added because averaging changes the noise distribution in a way that makes the subsequent VST less accurate.

Visual comparison of denoising results across four benchmark images. From left to right: original image, noisy image, BM3D-4 result, and BM3D-4$_{\sigma}$ result.

As seen in the visual comparisons, the low-pass preprocessing step (BM3D-4$_{\sigma}$) provides a significant quality jump, particularly in high-noise scenarios. It achieved up to a 0.94 dB improvement in PSNR for the 10-frame Lena dataset. Despite the increased complexity of BM3D-4, a CUDA implementation on an NVIDIA GeForce GTX 1070 can process a $256 \times 256 \times 5$ image stack in just 1.76 seconds.

Conclusion

Maximizing "inter-frame connectivity" is the key to effective multi-frame denoising. By allowing every pixel in every frame to participate in multiple 3D groups, BM3D-4 aggregates significantly more information than simpler averaging methods. Future work will focus on relaxing the assumption of perfect frame registration to handle motion in real-world video and exploring "Poisson-Gaussian" mixture noise models to better simulate electronic sensor read noise.

Summary updated

Abstract

The 3D block matching (BM3D) filter belongs to the state-of-the-art techniques for eliminating additive white Gaussian noise from single-frame images. There exist four multi-frame extensions of BM3D as of today. In this work, we combine these extensions with a variance stabilising transformation (VST) for eliminating Poisson noise. Our evaluation reveals that the extension which retains the original noise model of the noisy images and additionally has a comprehensive connectivity of 2D and temporal image information at both pixel and patch levels, gives the best results. Additionally, we find a surprising change in performance of one the four extensions due to the specific application of the VST. Finally, we also introduce a simple low-pass filtering as a preprocessing step for the best performing extension. This can give rise to a significant additional improvement of 0.94 dB in the output according to the peak signal to noise ratio.
BibTeX
@inproceedings{bodduna2019poisson,
  author = {Kireeti Bodduna and Joachim Weickert},
  title = {{Poisson Noise Removal Using Multi-Frame 3D Block Matching}},
  booktitle = {EUVIP},
  year = {2019},
  doi = {10.1109/euvip47703.2019.8946209},
}
Figure